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/types.h>
27 #include <linux/workqueue.h>
28 
29 #include "cpufreq_governor.h"
30 
31 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
32 {
33 	if (have_governor_per_policy())
34 		return dbs_data->cdata->attr_group_gov_pol;
35 	else
36 		return dbs_data->cdata->attr_group_gov_sys;
37 }
38 
39 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
40 {
41 	struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
42 	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
43 	struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
44 	struct cpufreq_policy *policy;
45 	unsigned int max_load = 0;
46 	unsigned int ignore_nice;
47 	unsigned int j;
48 
49 	if (dbs_data->cdata->governor == GOV_ONDEMAND)
50 		ignore_nice = od_tuners->ignore_nice_load;
51 	else
52 		ignore_nice = cs_tuners->ignore_nice_load;
53 
54 	policy = cdbs->cur_policy;
55 
56 	/* Get Absolute Load (in terms of freq for ondemand gov) */
57 	for_each_cpu(j, policy->cpus) {
58 		struct cpu_dbs_common_info *j_cdbs;
59 		u64 cur_wall_time, cur_idle_time;
60 		unsigned int idle_time, wall_time;
61 		unsigned int load;
62 		int io_busy = 0;
63 
64 		j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
65 
66 		/*
67 		 * For the purpose of ondemand, waiting for disk IO is
68 		 * an indication that you're performance critical, and
69 		 * not that the system is actually idle. So do not add
70 		 * the iowait time to the cpu idle time.
71 		 */
72 		if (dbs_data->cdata->governor == GOV_ONDEMAND)
73 			io_busy = od_tuners->io_is_busy;
74 		cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
75 
76 		wall_time = (unsigned int)
77 			(cur_wall_time - j_cdbs->prev_cpu_wall);
78 		j_cdbs->prev_cpu_wall = cur_wall_time;
79 
80 		idle_time = (unsigned int)
81 			(cur_idle_time - j_cdbs->prev_cpu_idle);
82 		j_cdbs->prev_cpu_idle = cur_idle_time;
83 
84 		if (ignore_nice) {
85 			u64 cur_nice;
86 			unsigned long cur_nice_jiffies;
87 
88 			cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
89 					 cdbs->prev_cpu_nice;
90 			/*
91 			 * Assumption: nice time between sampling periods will
92 			 * be less than 2^32 jiffies for 32 bit sys
93 			 */
94 			cur_nice_jiffies = (unsigned long)
95 					cputime64_to_jiffies64(cur_nice);
96 
97 			cdbs->prev_cpu_nice =
98 				kcpustat_cpu(j).cpustat[CPUTIME_NICE];
99 			idle_time += jiffies_to_usecs(cur_nice_jiffies);
100 		}
101 
102 		if (unlikely(!wall_time || wall_time < idle_time))
103 			continue;
104 
105 		load = 100 * (wall_time - idle_time) / wall_time;
106 
107 		if (dbs_data->cdata->governor == GOV_ONDEMAND) {
108 			int freq_avg = __cpufreq_driver_getavg(policy, j);
109 			if (freq_avg <= 0)
110 				freq_avg = policy->cur;
111 
112 			load *= freq_avg;
113 		}
114 
115 		if (load > max_load)
116 			max_load = load;
117 	}
118 
119 	dbs_data->cdata->gov_check_cpu(cpu, max_load);
120 }
121 EXPORT_SYMBOL_GPL(dbs_check_cpu);
122 
123 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
124 		unsigned int delay)
125 {
126 	struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
127 
128 	mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
129 }
130 
131 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
132 		unsigned int delay, bool all_cpus)
133 {
134 	int i;
135 
136 	if (!all_cpus) {
137 		__gov_queue_work(smp_processor_id(), dbs_data, delay);
138 	} else {
139 		for_each_cpu(i, policy->cpus)
140 			__gov_queue_work(i, dbs_data, delay);
141 	}
142 }
143 EXPORT_SYMBOL_GPL(gov_queue_work);
144 
145 static inline void gov_cancel_work(struct dbs_data *dbs_data,
146 		struct cpufreq_policy *policy)
147 {
148 	struct cpu_dbs_common_info *cdbs;
149 	int i;
150 
151 	for_each_cpu(i, policy->cpus) {
152 		cdbs = dbs_data->cdata->get_cpu_cdbs(i);
153 		cancel_delayed_work_sync(&cdbs->work);
154 	}
155 }
156 
157 /* Will return if we need to evaluate cpu load again or not */
158 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
159 		unsigned int sampling_rate)
160 {
161 	if (policy_is_shared(cdbs->cur_policy)) {
162 		ktime_t time_now = ktime_get();
163 		s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
164 
165 		/* Do nothing if we recently have sampled */
166 		if (delta_us < (s64)(sampling_rate / 2))
167 			return false;
168 		else
169 			cdbs->time_stamp = time_now;
170 	}
171 
172 	return true;
173 }
174 EXPORT_SYMBOL_GPL(need_load_eval);
175 
176 static void set_sampling_rate(struct dbs_data *dbs_data,
177 		unsigned int sampling_rate)
178 {
179 	if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
180 		struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
181 		cs_tuners->sampling_rate = sampling_rate;
182 	} else {
183 		struct od_dbs_tuners *od_tuners = dbs_data->tuners;
184 		od_tuners->sampling_rate = sampling_rate;
185 	}
186 }
187 
188 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
189 		struct common_dbs_data *cdata, unsigned int event)
190 {
191 	struct dbs_data *dbs_data;
192 	struct od_cpu_dbs_info_s *od_dbs_info = NULL;
193 	struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
194 	struct od_ops *od_ops = NULL;
195 	struct od_dbs_tuners *od_tuners = NULL;
196 	struct cs_dbs_tuners *cs_tuners = NULL;
197 	struct cpu_dbs_common_info *cpu_cdbs;
198 	unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
199 	int io_busy = 0;
200 	int rc;
201 
202 	if (have_governor_per_policy())
203 		dbs_data = policy->governor_data;
204 	else
205 		dbs_data = cdata->gdbs_data;
206 
207 	WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
208 
209 	switch (event) {
210 	case CPUFREQ_GOV_POLICY_INIT:
211 		if (have_governor_per_policy()) {
212 			WARN_ON(dbs_data);
213 		} else if (dbs_data) {
214 			dbs_data->usage_count++;
215 			policy->governor_data = dbs_data;
216 			return 0;
217 		}
218 
219 		dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
220 		if (!dbs_data) {
221 			pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
222 			return -ENOMEM;
223 		}
224 
225 		dbs_data->cdata = cdata;
226 		dbs_data->usage_count = 1;
227 		rc = cdata->init(dbs_data);
228 		if (rc) {
229 			pr_err("%s: POLICY_INIT: init() failed\n", __func__);
230 			kfree(dbs_data);
231 			return rc;
232 		}
233 
234 		if (!have_governor_per_policy())
235 			WARN_ON(cpufreq_get_global_kobject());
236 
237 		rc = sysfs_create_group(get_governor_parent_kobj(policy),
238 				get_sysfs_attr(dbs_data));
239 		if (rc) {
240 			cdata->exit(dbs_data);
241 			kfree(dbs_data);
242 			return rc;
243 		}
244 
245 		policy->governor_data = dbs_data;
246 
247 		/* policy latency is in nS. Convert it to uS first */
248 		latency = policy->cpuinfo.transition_latency / 1000;
249 		if (latency == 0)
250 			latency = 1;
251 
252 		/* Bring kernel and HW constraints together */
253 		dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
254 				MIN_LATENCY_MULTIPLIER * latency);
255 		set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
256 					latency * LATENCY_MULTIPLIER));
257 
258 		if ((cdata->governor == GOV_CONSERVATIVE) &&
259 				(!policy->governor->initialized)) {
260 			struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
261 
262 			cpufreq_register_notifier(cs_ops->notifier_block,
263 					CPUFREQ_TRANSITION_NOTIFIER);
264 		}
265 
266 		if (!have_governor_per_policy())
267 			cdata->gdbs_data = dbs_data;
268 
269 		return 0;
270 	case CPUFREQ_GOV_POLICY_EXIT:
271 		if (!--dbs_data->usage_count) {
272 			sysfs_remove_group(get_governor_parent_kobj(policy),
273 					get_sysfs_attr(dbs_data));
274 
275 			if (!have_governor_per_policy())
276 				cpufreq_put_global_kobject();
277 
278 			if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
279 				(policy->governor->initialized == 1)) {
280 				struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
281 
282 				cpufreq_unregister_notifier(cs_ops->notifier_block,
283 						CPUFREQ_TRANSITION_NOTIFIER);
284 			}
285 
286 			cdata->exit(dbs_data);
287 			kfree(dbs_data);
288 			cdata->gdbs_data = NULL;
289 		}
290 
291 		policy->governor_data = NULL;
292 		return 0;
293 	}
294 
295 	cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
296 
297 	if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
298 		cs_tuners = dbs_data->tuners;
299 		cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
300 		sampling_rate = cs_tuners->sampling_rate;
301 		ignore_nice = cs_tuners->ignore_nice_load;
302 	} else {
303 		od_tuners = dbs_data->tuners;
304 		od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
305 		sampling_rate = od_tuners->sampling_rate;
306 		ignore_nice = od_tuners->ignore_nice_load;
307 		od_ops = dbs_data->cdata->gov_ops;
308 		io_busy = od_tuners->io_is_busy;
309 	}
310 
311 	switch (event) {
312 	case CPUFREQ_GOV_START:
313 		if (!policy->cur)
314 			return -EINVAL;
315 
316 		mutex_lock(&dbs_data->mutex);
317 
318 		for_each_cpu(j, policy->cpus) {
319 			struct cpu_dbs_common_info *j_cdbs =
320 				dbs_data->cdata->get_cpu_cdbs(j);
321 
322 			j_cdbs->cpu = j;
323 			j_cdbs->cur_policy = policy;
324 			j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
325 					       &j_cdbs->prev_cpu_wall, io_busy);
326 			if (ignore_nice)
327 				j_cdbs->prev_cpu_nice =
328 					kcpustat_cpu(j).cpustat[CPUTIME_NICE];
329 
330 			mutex_init(&j_cdbs->timer_mutex);
331 			INIT_DEFERRABLE_WORK(&j_cdbs->work,
332 					     dbs_data->cdata->gov_dbs_timer);
333 		}
334 
335 		/*
336 		 * conservative does not implement micro like ondemand
337 		 * governor, thus we are bound to jiffes/HZ
338 		 */
339 		if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
340 			cs_dbs_info->down_skip = 0;
341 			cs_dbs_info->enable = 1;
342 			cs_dbs_info->requested_freq = policy->cur;
343 		} else {
344 			od_dbs_info->rate_mult = 1;
345 			od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
346 			od_ops->powersave_bias_init_cpu(cpu);
347 		}
348 
349 		mutex_unlock(&dbs_data->mutex);
350 
351 		/* Initiate timer time stamp */
352 		cpu_cdbs->time_stamp = ktime_get();
353 
354 		gov_queue_work(dbs_data, policy,
355 				delay_for_sampling_rate(sampling_rate), true);
356 		break;
357 
358 	case CPUFREQ_GOV_STOP:
359 		if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
360 			cs_dbs_info->enable = 0;
361 
362 		gov_cancel_work(dbs_data, policy);
363 
364 		mutex_lock(&dbs_data->mutex);
365 		mutex_destroy(&cpu_cdbs->timer_mutex);
366 		cpu_cdbs->cur_policy = NULL;
367 
368 		mutex_unlock(&dbs_data->mutex);
369 
370 		break;
371 
372 	case CPUFREQ_GOV_LIMITS:
373 		mutex_lock(&cpu_cdbs->timer_mutex);
374 		if (policy->max < cpu_cdbs->cur_policy->cur)
375 			__cpufreq_driver_target(cpu_cdbs->cur_policy,
376 					policy->max, CPUFREQ_RELATION_H);
377 		else if (policy->min > cpu_cdbs->cur_policy->cur)
378 			__cpufreq_driver_target(cpu_cdbs->cur_policy,
379 					policy->min, CPUFREQ_RELATION_L);
380 		dbs_check_cpu(dbs_data, cpu);
381 		mutex_unlock(&cpu_cdbs->timer_mutex);
382 		break;
383 	}
384 	return 0;
385 }
386 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);
387