1 /*
2  *  drivers/cpufreq/cpufreq_ondemand.c
3  *
4  *  Copyright (C)  2001 Russell King
5  *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6  *                      Jun Nakajima <jun.nakajima@intel.com>
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 as
10  * published by the Free Software Foundation.
11  */
12 
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 
15 #include <linux/cpu.h>
16 #include <linux/percpu-defs.h>
17 #include <linux/slab.h>
18 #include <linux/tick.h>
19 #include <linux/sched/cpufreq.h>
20 
21 #include "cpufreq_ondemand.h"
22 
23 /* On-demand governor macros */
24 #define DEF_FREQUENCY_UP_THRESHOLD		(80)
25 #define DEF_SAMPLING_DOWN_FACTOR		(1)
26 #define MAX_SAMPLING_DOWN_FACTOR		(100000)
27 #define MICRO_FREQUENCY_UP_THRESHOLD		(95)
28 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE		(10000)
29 #define MIN_FREQUENCY_UP_THRESHOLD		(1)
30 #define MAX_FREQUENCY_UP_THRESHOLD		(100)
31 
32 static struct od_ops od_ops;
33 
34 static unsigned int default_powersave_bias;
35 
36 /*
37  * Not all CPUs want IO time to be accounted as busy; this depends on how
38  * efficient idling at a higher frequency/voltage is.
39  * Pavel Machek says this is not so for various generations of AMD and old
40  * Intel systems.
41  * Mike Chan (android.com) claims this is also not true for ARM.
42  * Because of this, whitelist specific known (series) of CPUs by default, and
43  * leave all others up to the user.
44  */
45 static int should_io_be_busy(void)
46 {
47 #if defined(CONFIG_X86)
48 	/*
49 	 * For Intel, Core 2 (model 15) and later have an efficient idle.
50 	 */
51 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
52 			boot_cpu_data.x86 == 6 &&
53 			boot_cpu_data.x86_model >= 15)
54 		return 1;
55 #endif
56 	return 0;
57 }
58 
59 /*
60  * Find right freq to be set now with powersave_bias on.
61  * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
62  * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
63  */
64 static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
65 		unsigned int freq_next, unsigned int relation)
66 {
67 	unsigned int freq_req, freq_reduc, freq_avg;
68 	unsigned int freq_hi, freq_lo;
69 	unsigned int index;
70 	unsigned int delay_hi_us;
71 	struct policy_dbs_info *policy_dbs = policy->governor_data;
72 	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
73 	struct dbs_data *dbs_data = policy_dbs->dbs_data;
74 	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
75 	struct cpufreq_frequency_table *freq_table = policy->freq_table;
76 
77 	if (!freq_table) {
78 		dbs_info->freq_lo = 0;
79 		dbs_info->freq_lo_delay_us = 0;
80 		return freq_next;
81 	}
82 
83 	index = cpufreq_frequency_table_target(policy, freq_next, relation);
84 	freq_req = freq_table[index].frequency;
85 	freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
86 	freq_avg = freq_req - freq_reduc;
87 
88 	/* Find freq bounds for freq_avg in freq_table */
89 	index = cpufreq_table_find_index_h(policy, freq_avg);
90 	freq_lo = freq_table[index].frequency;
91 	index = cpufreq_table_find_index_l(policy, freq_avg);
92 	freq_hi = freq_table[index].frequency;
93 
94 	/* Find out how long we have to be in hi and lo freqs */
95 	if (freq_hi == freq_lo) {
96 		dbs_info->freq_lo = 0;
97 		dbs_info->freq_lo_delay_us = 0;
98 		return freq_lo;
99 	}
100 	delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
101 	delay_hi_us += (freq_hi - freq_lo) / 2;
102 	delay_hi_us /= freq_hi - freq_lo;
103 	dbs_info->freq_hi_delay_us = delay_hi_us;
104 	dbs_info->freq_lo = freq_lo;
105 	dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
106 	return freq_hi;
107 }
108 
109 static void ondemand_powersave_bias_init(struct cpufreq_policy *policy)
110 {
111 	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
112 
113 	dbs_info->freq_lo = 0;
114 }
115 
116 static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
117 {
118 	struct policy_dbs_info *policy_dbs = policy->governor_data;
119 	struct dbs_data *dbs_data = policy_dbs->dbs_data;
120 	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
121 
122 	if (od_tuners->powersave_bias)
123 		freq = od_ops.powersave_bias_target(policy, freq,
124 				CPUFREQ_RELATION_H);
125 	else if (policy->cur == policy->max)
126 		return;
127 
128 	__cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
129 			CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
130 }
131 
132 /*
133  * Every sampling_rate, we check, if current idle time is less than 20%
134  * (default), then we try to increase frequency. Else, we adjust the frequency
135  * proportional to load.
136  */
137 static void od_update(struct cpufreq_policy *policy)
138 {
139 	struct policy_dbs_info *policy_dbs = policy->governor_data;
140 	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
141 	struct dbs_data *dbs_data = policy_dbs->dbs_data;
142 	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
143 	unsigned int load = dbs_update(policy);
144 
145 	dbs_info->freq_lo = 0;
146 
147 	/* Check for frequency increase */
148 	if (load > dbs_data->up_threshold) {
149 		/* If switching to max speed, apply sampling_down_factor */
150 		if (policy->cur < policy->max)
151 			policy_dbs->rate_mult = dbs_data->sampling_down_factor;
152 		dbs_freq_increase(policy, policy->max);
153 	} else {
154 		/* Calculate the next frequency proportional to load */
155 		unsigned int freq_next, min_f, max_f;
156 
157 		min_f = policy->cpuinfo.min_freq;
158 		max_f = policy->cpuinfo.max_freq;
159 		freq_next = min_f + load * (max_f - min_f) / 100;
160 
161 		/* No longer fully busy, reset rate_mult */
162 		policy_dbs->rate_mult = 1;
163 
164 		if (od_tuners->powersave_bias)
165 			freq_next = od_ops.powersave_bias_target(policy,
166 								 freq_next,
167 								 CPUFREQ_RELATION_L);
168 
169 		__cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
170 	}
171 }
172 
173 static unsigned int od_dbs_update(struct cpufreq_policy *policy)
174 {
175 	struct policy_dbs_info *policy_dbs = policy->governor_data;
176 	struct dbs_data *dbs_data = policy_dbs->dbs_data;
177 	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
178 	int sample_type = dbs_info->sample_type;
179 
180 	/* Common NORMAL_SAMPLE setup */
181 	dbs_info->sample_type = OD_NORMAL_SAMPLE;
182 	/*
183 	 * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
184 	 * it then.
185 	 */
186 	if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
187 		__cpufreq_driver_target(policy, dbs_info->freq_lo,
188 					CPUFREQ_RELATION_H);
189 		return dbs_info->freq_lo_delay_us;
190 	}
191 
192 	od_update(policy);
193 
194 	if (dbs_info->freq_lo) {
195 		/* Setup SUB_SAMPLE */
196 		dbs_info->sample_type = OD_SUB_SAMPLE;
197 		return dbs_info->freq_hi_delay_us;
198 	}
199 
200 	return dbs_data->sampling_rate * policy_dbs->rate_mult;
201 }
202 
203 /************************** sysfs interface ************************/
204 static struct dbs_governor od_dbs_gov;
205 
206 static ssize_t store_io_is_busy(struct gov_attr_set *attr_set, const char *buf,
207 				size_t count)
208 {
209 	struct dbs_data *dbs_data = to_dbs_data(attr_set);
210 	unsigned int input;
211 	int ret;
212 
213 	ret = sscanf(buf, "%u", &input);
214 	if (ret != 1)
215 		return -EINVAL;
216 	dbs_data->io_is_busy = !!input;
217 
218 	/* we need to re-evaluate prev_cpu_idle */
219 	gov_update_cpu_data(dbs_data);
220 
221 	return count;
222 }
223 
224 static ssize_t store_up_threshold(struct gov_attr_set *attr_set,
225 				  const char *buf, size_t count)
226 {
227 	struct dbs_data *dbs_data = to_dbs_data(attr_set);
228 	unsigned int input;
229 	int ret;
230 	ret = sscanf(buf, "%u", &input);
231 
232 	if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
233 			input < MIN_FREQUENCY_UP_THRESHOLD) {
234 		return -EINVAL;
235 	}
236 
237 	dbs_data->up_threshold = input;
238 	return count;
239 }
240 
241 static ssize_t store_sampling_down_factor(struct gov_attr_set *attr_set,
242 					  const char *buf, size_t count)
243 {
244 	struct dbs_data *dbs_data = to_dbs_data(attr_set);
245 	struct policy_dbs_info *policy_dbs;
246 	unsigned int input;
247 	int ret;
248 	ret = sscanf(buf, "%u", &input);
249 
250 	if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
251 		return -EINVAL;
252 
253 	dbs_data->sampling_down_factor = input;
254 
255 	/* Reset down sampling multiplier in case it was active */
256 	list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
257 		/*
258 		 * Doing this without locking might lead to using different
259 		 * rate_mult values in od_update() and od_dbs_update().
260 		 */
261 		mutex_lock(&policy_dbs->update_mutex);
262 		policy_dbs->rate_mult = 1;
263 		mutex_unlock(&policy_dbs->update_mutex);
264 	}
265 
266 	return count;
267 }
268 
269 static ssize_t store_ignore_nice_load(struct gov_attr_set *attr_set,
270 				      const char *buf, size_t count)
271 {
272 	struct dbs_data *dbs_data = to_dbs_data(attr_set);
273 	unsigned int input;
274 	int ret;
275 
276 	ret = sscanf(buf, "%u", &input);
277 	if (ret != 1)
278 		return -EINVAL;
279 
280 	if (input > 1)
281 		input = 1;
282 
283 	if (input == dbs_data->ignore_nice_load) { /* nothing to do */
284 		return count;
285 	}
286 	dbs_data->ignore_nice_load = input;
287 
288 	/* we need to re-evaluate prev_cpu_idle */
289 	gov_update_cpu_data(dbs_data);
290 
291 	return count;
292 }
293 
294 static ssize_t store_powersave_bias(struct gov_attr_set *attr_set,
295 				    const char *buf, size_t count)
296 {
297 	struct dbs_data *dbs_data = to_dbs_data(attr_set);
298 	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
299 	struct policy_dbs_info *policy_dbs;
300 	unsigned int input;
301 	int ret;
302 	ret = sscanf(buf, "%u", &input);
303 
304 	if (ret != 1)
305 		return -EINVAL;
306 
307 	if (input > 1000)
308 		input = 1000;
309 
310 	od_tuners->powersave_bias = input;
311 
312 	list_for_each_entry(policy_dbs, &attr_set->policy_list, list)
313 		ondemand_powersave_bias_init(policy_dbs->policy);
314 
315 	return count;
316 }
317 
318 gov_show_one_common(sampling_rate);
319 gov_show_one_common(up_threshold);
320 gov_show_one_common(sampling_down_factor);
321 gov_show_one_common(ignore_nice_load);
322 gov_show_one_common(min_sampling_rate);
323 gov_show_one_common(io_is_busy);
324 gov_show_one(od, powersave_bias);
325 
326 gov_attr_rw(sampling_rate);
327 gov_attr_rw(io_is_busy);
328 gov_attr_rw(up_threshold);
329 gov_attr_rw(sampling_down_factor);
330 gov_attr_rw(ignore_nice_load);
331 gov_attr_rw(powersave_bias);
332 gov_attr_ro(min_sampling_rate);
333 
334 static struct attribute *od_attributes[] = {
335 	&min_sampling_rate.attr,
336 	&sampling_rate.attr,
337 	&up_threshold.attr,
338 	&sampling_down_factor.attr,
339 	&ignore_nice_load.attr,
340 	&powersave_bias.attr,
341 	&io_is_busy.attr,
342 	NULL
343 };
344 
345 /************************** sysfs end ************************/
346 
347 static struct policy_dbs_info *od_alloc(void)
348 {
349 	struct od_policy_dbs_info *dbs_info;
350 
351 	dbs_info = kzalloc(sizeof(*dbs_info), GFP_KERNEL);
352 	return dbs_info ? &dbs_info->policy_dbs : NULL;
353 }
354 
355 static void od_free(struct policy_dbs_info *policy_dbs)
356 {
357 	kfree(to_dbs_info(policy_dbs));
358 }
359 
360 static int od_init(struct dbs_data *dbs_data)
361 {
362 	struct od_dbs_tuners *tuners;
363 	u64 idle_time;
364 	int cpu;
365 
366 	tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
367 	if (!tuners)
368 		return -ENOMEM;
369 
370 	cpu = get_cpu();
371 	idle_time = get_cpu_idle_time_us(cpu, NULL);
372 	put_cpu();
373 	if (idle_time != -1ULL) {
374 		/* Idle micro accounting is supported. Use finer thresholds */
375 		dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
376 		/*
377 		 * In nohz/micro accounting case we set the minimum frequency
378 		 * not depending on HZ, but fixed (very low).
379 		*/
380 		dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
381 	} else {
382 		dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
383 
384 		/* For correct statistics, we need 10 ticks for each measure */
385 		dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
386 			jiffies_to_usecs(10);
387 	}
388 
389 	dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
390 	dbs_data->ignore_nice_load = 0;
391 	tuners->powersave_bias = default_powersave_bias;
392 	dbs_data->io_is_busy = should_io_be_busy();
393 
394 	dbs_data->tuners = tuners;
395 	return 0;
396 }
397 
398 static void od_exit(struct dbs_data *dbs_data)
399 {
400 	kfree(dbs_data->tuners);
401 }
402 
403 static void od_start(struct cpufreq_policy *policy)
404 {
405 	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
406 
407 	dbs_info->sample_type = OD_NORMAL_SAMPLE;
408 	ondemand_powersave_bias_init(policy);
409 }
410 
411 static struct od_ops od_ops = {
412 	.powersave_bias_target = generic_powersave_bias_target,
413 };
414 
415 static struct dbs_governor od_dbs_gov = {
416 	.gov = CPUFREQ_DBS_GOVERNOR_INITIALIZER("ondemand"),
417 	.kobj_type = { .default_attrs = od_attributes },
418 	.gov_dbs_update = od_dbs_update,
419 	.alloc = od_alloc,
420 	.free = od_free,
421 	.init = od_init,
422 	.exit = od_exit,
423 	.start = od_start,
424 };
425 
426 #define CPU_FREQ_GOV_ONDEMAND	(&od_dbs_gov.gov)
427 
428 static void od_set_powersave_bias(unsigned int powersave_bias)
429 {
430 	unsigned int cpu;
431 	cpumask_t done;
432 
433 	default_powersave_bias = powersave_bias;
434 	cpumask_clear(&done);
435 
436 	get_online_cpus();
437 	for_each_online_cpu(cpu) {
438 		struct cpufreq_policy *policy;
439 		struct policy_dbs_info *policy_dbs;
440 		struct dbs_data *dbs_data;
441 		struct od_dbs_tuners *od_tuners;
442 
443 		if (cpumask_test_cpu(cpu, &done))
444 			continue;
445 
446 		policy = cpufreq_cpu_get_raw(cpu);
447 		if (!policy || policy->governor != CPU_FREQ_GOV_ONDEMAND)
448 			continue;
449 
450 		policy_dbs = policy->governor_data;
451 		if (!policy_dbs)
452 			continue;
453 
454 		cpumask_or(&done, &done, policy->cpus);
455 
456 		dbs_data = policy_dbs->dbs_data;
457 		od_tuners = dbs_data->tuners;
458 		od_tuners->powersave_bias = default_powersave_bias;
459 	}
460 	put_online_cpus();
461 }
462 
463 void od_register_powersave_bias_handler(unsigned int (*f)
464 		(struct cpufreq_policy *, unsigned int, unsigned int),
465 		unsigned int powersave_bias)
466 {
467 	od_ops.powersave_bias_target = f;
468 	od_set_powersave_bias(powersave_bias);
469 }
470 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
471 
472 void od_unregister_powersave_bias_handler(void)
473 {
474 	od_ops.powersave_bias_target = generic_powersave_bias_target;
475 	od_set_powersave_bias(0);
476 }
477 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
478 
479 static int __init cpufreq_gov_dbs_init(void)
480 {
481 	return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
482 }
483 
484 static void __exit cpufreq_gov_dbs_exit(void)
485 {
486 	cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
487 }
488 
489 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
490 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
491 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
492 	"Low Latency Frequency Transition capable processors");
493 MODULE_LICENSE("GPL");
494 
495 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
496 struct cpufreq_governor *cpufreq_default_governor(void)
497 {
498 	return CPU_FREQ_GOV_ONDEMAND;
499 }
500 
501 fs_initcall(cpufreq_gov_dbs_init);
502 #else
503 module_init(cpufreq_gov_dbs_init);
504 #endif
505 module_exit(cpufreq_gov_dbs_exit);
506