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 "cpufreq_governor.h"
20
21/* On-demand governor macros */
22#define DEF_FREQUENCY_UP_THRESHOLD		(80)
23#define DEF_SAMPLING_DOWN_FACTOR		(1)
24#define MAX_SAMPLING_DOWN_FACTOR		(100000)
25#define MICRO_FREQUENCY_UP_THRESHOLD		(95)
26#define MICRO_FREQUENCY_MIN_SAMPLE_RATE		(10000)
27#define MIN_FREQUENCY_UP_THRESHOLD		(11)
28#define MAX_FREQUENCY_UP_THRESHOLD		(100)
29
30static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
31
32static struct od_ops od_ops;
33
34#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
35static struct cpufreq_governor cpufreq_gov_ondemand;
36#endif
37
38static unsigned int default_powersave_bias;
39
40static void ondemand_powersave_bias_init_cpu(int cpu)
41{
42	struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
43
44	dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
45	dbs_info->freq_lo = 0;
46}
47
48/*
49 * Not all CPUs want IO time to be accounted as busy; this depends on how
50 * efficient idling at a higher frequency/voltage is.
51 * Pavel Machek says this is not so for various generations of AMD and old
52 * Intel systems.
53 * Mike Chan (android.com) claims this is also not true for ARM.
54 * Because of this, whitelist specific known (series) of CPUs by default, and
55 * leave all others up to the user.
56 */
57static int should_io_be_busy(void)
58{
59#if defined(CONFIG_X86)
60	/*
61	 * For Intel, Core 2 (model 15) and later have an efficient idle.
62	 */
63	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
64			boot_cpu_data.x86 == 6 &&
65			boot_cpu_data.x86_model >= 15)
66		return 1;
67#endif
68	return 0;
69}
70
71/*
72 * Find right freq to be set now with powersave_bias on.
73 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
74 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
75 */
76static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
77		unsigned int freq_next, unsigned int relation)
78{
79	unsigned int freq_req, freq_reduc, freq_avg;
80	unsigned int freq_hi, freq_lo;
81	unsigned int index = 0;
82	unsigned int jiffies_total, jiffies_hi, jiffies_lo;
83	struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
84						   policy->cpu);
85	struct dbs_data *dbs_data = policy->governor_data;
86	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
87
88	if (!dbs_info->freq_table) {
89		dbs_info->freq_lo = 0;
90		dbs_info->freq_lo_jiffies = 0;
91		return freq_next;
92	}
93
94	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
95			relation, &index);
96	freq_req = dbs_info->freq_table[index].frequency;
97	freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
98	freq_avg = freq_req - freq_reduc;
99
100	/* Find freq bounds for freq_avg in freq_table */
101	index = 0;
102	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
103			CPUFREQ_RELATION_H, &index);
104	freq_lo = dbs_info->freq_table[index].frequency;
105	index = 0;
106	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
107			CPUFREQ_RELATION_L, &index);
108	freq_hi = dbs_info->freq_table[index].frequency;
109
110	/* Find out how long we have to be in hi and lo freqs */
111	if (freq_hi == freq_lo) {
112		dbs_info->freq_lo = 0;
113		dbs_info->freq_lo_jiffies = 0;
114		return freq_lo;
115	}
116	jiffies_total = usecs_to_jiffies(od_tuners->sampling_rate);
117	jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
118	jiffies_hi += ((freq_hi - freq_lo) / 2);
119	jiffies_hi /= (freq_hi - freq_lo);
120	jiffies_lo = jiffies_total - jiffies_hi;
121	dbs_info->freq_lo = freq_lo;
122	dbs_info->freq_lo_jiffies = jiffies_lo;
123	dbs_info->freq_hi_jiffies = jiffies_hi;
124	return freq_hi;
125}
126
127static void ondemand_powersave_bias_init(void)
128{
129	int i;
130	for_each_online_cpu(i) {
131		ondemand_powersave_bias_init_cpu(i);
132	}
133}
134
135static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
136{
137	struct dbs_data *dbs_data = policy->governor_data;
138	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
139
140	if (od_tuners->powersave_bias)
141		freq = od_ops.powersave_bias_target(policy, freq,
142				CPUFREQ_RELATION_H);
143	else if (policy->cur == policy->max)
144		return;
145
146	__cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
147			CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
148}
149
150/*
151 * Every sampling_rate, we check, if current idle time is less than 20%
152 * (default), then we try to increase frequency. Else, we adjust the frequency
153 * proportional to load.
154 */
155static void od_check_cpu(int cpu, unsigned int load)
156{
157	struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
158	struct cpufreq_policy *policy = dbs_info->cdbs.shared->policy;
159	struct dbs_data *dbs_data = policy->governor_data;
160	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
161
162	dbs_info->freq_lo = 0;
163
164	/* Check for frequency increase */
165	if (load > od_tuners->up_threshold) {
166		/* If switching to max speed, apply sampling_down_factor */
167		if (policy->cur < policy->max)
168			dbs_info->rate_mult =
169				od_tuners->sampling_down_factor;
170		dbs_freq_increase(policy, policy->max);
171	} else {
172		/* Calculate the next frequency proportional to load */
173		unsigned int freq_next, min_f, max_f;
174
175		min_f = policy->cpuinfo.min_freq;
176		max_f = policy->cpuinfo.max_freq;
177		freq_next = min_f + load * (max_f - min_f) / 100;
178
179		/* No longer fully busy, reset rate_mult */
180		dbs_info->rate_mult = 1;
181
182		if (!od_tuners->powersave_bias) {
183			__cpufreq_driver_target(policy, freq_next,
184					CPUFREQ_RELATION_C);
185			return;
186		}
187
188		freq_next = od_ops.powersave_bias_target(policy, freq_next,
189					CPUFREQ_RELATION_L);
190		__cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
191	}
192}
193
194static unsigned int od_dbs_timer(struct cpu_dbs_info *cdbs,
195				 struct dbs_data *dbs_data, bool modify_all)
196{
197	struct cpufreq_policy *policy = cdbs->shared->policy;
198	unsigned int cpu = policy->cpu;
199	struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
200			cpu);
201	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
202	int delay = 0, sample_type = dbs_info->sample_type;
203
204	if (!modify_all)
205		goto max_delay;
206
207	/* Common NORMAL_SAMPLE setup */
208	dbs_info->sample_type = OD_NORMAL_SAMPLE;
209	if (sample_type == OD_SUB_SAMPLE) {
210		delay = dbs_info->freq_lo_jiffies;
211		__cpufreq_driver_target(policy, dbs_info->freq_lo,
212					CPUFREQ_RELATION_H);
213	} else {
214		dbs_check_cpu(dbs_data, cpu);
215		if (dbs_info->freq_lo) {
216			/* Setup timer for SUB_SAMPLE */
217			dbs_info->sample_type = OD_SUB_SAMPLE;
218			delay = dbs_info->freq_hi_jiffies;
219		}
220	}
221
222max_delay:
223	if (!delay)
224		delay = delay_for_sampling_rate(od_tuners->sampling_rate
225				* dbs_info->rate_mult);
226
227	return delay;
228}
229
230/************************** sysfs interface ************************/
231static struct common_dbs_data od_dbs_cdata;
232
233/**
234 * update_sampling_rate - update sampling rate effective immediately if needed.
235 * @new_rate: new sampling rate
236 *
237 * If new rate is smaller than the old, simply updating
238 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
239 * original sampling_rate was 1 second and the requested new sampling rate is 10
240 * ms because the user needs immediate reaction from ondemand governor, but not
241 * sure if higher frequency will be required or not, then, the governor may
242 * change the sampling rate too late; up to 1 second later. Thus, if we are
243 * reducing the sampling rate, we need to make the new value effective
244 * immediately.
245 */
246static void update_sampling_rate(struct dbs_data *dbs_data,
247		unsigned int new_rate)
248{
249	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
250	int cpu;
251
252	od_tuners->sampling_rate = new_rate = max(new_rate,
253			dbs_data->min_sampling_rate);
254
255	for_each_online_cpu(cpu) {
256		struct cpufreq_policy *policy;
257		struct od_cpu_dbs_info_s *dbs_info;
258		unsigned long next_sampling, appointed_at;
259
260		policy = cpufreq_cpu_get(cpu);
261		if (!policy)
262			continue;
263		if (policy->governor != &cpufreq_gov_ondemand) {
264			cpufreq_cpu_put(policy);
265			continue;
266		}
267		dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
268		cpufreq_cpu_put(policy);
269
270		if (!delayed_work_pending(&dbs_info->cdbs.dwork))
271			continue;
272
273		next_sampling = jiffies + usecs_to_jiffies(new_rate);
274		appointed_at = dbs_info->cdbs.dwork.timer.expires;
275
276		if (time_before(next_sampling, appointed_at)) {
277			cancel_delayed_work_sync(&dbs_info->cdbs.dwork);
278
279			gov_queue_work(dbs_data, policy,
280				       usecs_to_jiffies(new_rate), true);
281
282		}
283	}
284}
285
286static ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,
287		size_t count)
288{
289	unsigned int input;
290	int ret;
291	ret = sscanf(buf, "%u", &input);
292	if (ret != 1)
293		return -EINVAL;
294
295	update_sampling_rate(dbs_data, input);
296	return count;
297}
298
299static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf,
300		size_t count)
301{
302	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
303	unsigned int input;
304	int ret;
305	unsigned int j;
306
307	ret = sscanf(buf, "%u", &input);
308	if (ret != 1)
309		return -EINVAL;
310	od_tuners->io_is_busy = !!input;
311
312	/* we need to re-evaluate prev_cpu_idle */
313	for_each_online_cpu(j) {
314		struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
315									j);
316		dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
317			&dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
318	}
319	return count;
320}
321
322static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf,
323		size_t count)
324{
325	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
326	unsigned int input;
327	int ret;
328	ret = sscanf(buf, "%u", &input);
329
330	if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
331			input < MIN_FREQUENCY_UP_THRESHOLD) {
332		return -EINVAL;
333	}
334
335	od_tuners->up_threshold = input;
336	return count;
337}
338
339static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data,
340		const char *buf, size_t count)
341{
342	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
343	unsigned int input, j;
344	int ret;
345	ret = sscanf(buf, "%u", &input);
346
347	if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
348		return -EINVAL;
349	od_tuners->sampling_down_factor = input;
350
351	/* Reset down sampling multiplier in case it was active */
352	for_each_online_cpu(j) {
353		struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
354				j);
355		dbs_info->rate_mult = 1;
356	}
357	return count;
358}
359
360static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data,
361		const char *buf, size_t count)
362{
363	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
364	unsigned int input;
365	int ret;
366
367	unsigned int j;
368
369	ret = sscanf(buf, "%u", &input);
370	if (ret != 1)
371		return -EINVAL;
372
373	if (input > 1)
374		input = 1;
375
376	if (input == od_tuners->ignore_nice_load) { /* nothing to do */
377		return count;
378	}
379	od_tuners->ignore_nice_load = input;
380
381	/* we need to re-evaluate prev_cpu_idle */
382	for_each_online_cpu(j) {
383		struct od_cpu_dbs_info_s *dbs_info;
384		dbs_info = &per_cpu(od_cpu_dbs_info, j);
385		dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
386			&dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
387		if (od_tuners->ignore_nice_load)
388			dbs_info->cdbs.prev_cpu_nice =
389				kcpustat_cpu(j).cpustat[CPUTIME_NICE];
390
391	}
392	return count;
393}
394
395static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
396		size_t count)
397{
398	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
399	unsigned int input;
400	int ret;
401	ret = sscanf(buf, "%u", &input);
402
403	if (ret != 1)
404		return -EINVAL;
405
406	if (input > 1000)
407		input = 1000;
408
409	od_tuners->powersave_bias = input;
410	ondemand_powersave_bias_init();
411	return count;
412}
413
414show_store_one(od, sampling_rate);
415show_store_one(od, io_is_busy);
416show_store_one(od, up_threshold);
417show_store_one(od, sampling_down_factor);
418show_store_one(od, ignore_nice_load);
419show_store_one(od, powersave_bias);
420declare_show_sampling_rate_min(od);
421
422gov_sys_pol_attr_rw(sampling_rate);
423gov_sys_pol_attr_rw(io_is_busy);
424gov_sys_pol_attr_rw(up_threshold);
425gov_sys_pol_attr_rw(sampling_down_factor);
426gov_sys_pol_attr_rw(ignore_nice_load);
427gov_sys_pol_attr_rw(powersave_bias);
428gov_sys_pol_attr_ro(sampling_rate_min);
429
430static struct attribute *dbs_attributes_gov_sys[] = {
431	&sampling_rate_min_gov_sys.attr,
432	&sampling_rate_gov_sys.attr,
433	&up_threshold_gov_sys.attr,
434	&sampling_down_factor_gov_sys.attr,
435	&ignore_nice_load_gov_sys.attr,
436	&powersave_bias_gov_sys.attr,
437	&io_is_busy_gov_sys.attr,
438	NULL
439};
440
441static struct attribute_group od_attr_group_gov_sys = {
442	.attrs = dbs_attributes_gov_sys,
443	.name = "ondemand",
444};
445
446static struct attribute *dbs_attributes_gov_pol[] = {
447	&sampling_rate_min_gov_pol.attr,
448	&sampling_rate_gov_pol.attr,
449	&up_threshold_gov_pol.attr,
450	&sampling_down_factor_gov_pol.attr,
451	&ignore_nice_load_gov_pol.attr,
452	&powersave_bias_gov_pol.attr,
453	&io_is_busy_gov_pol.attr,
454	NULL
455};
456
457static struct attribute_group od_attr_group_gov_pol = {
458	.attrs = dbs_attributes_gov_pol,
459	.name = "ondemand",
460};
461
462/************************** sysfs end ************************/
463
464static int od_init(struct dbs_data *dbs_data, bool notify)
465{
466	struct od_dbs_tuners *tuners;
467	u64 idle_time;
468	int cpu;
469
470	tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
471	if (!tuners) {
472		pr_err("%s: kzalloc failed\n", __func__);
473		return -ENOMEM;
474	}
475
476	cpu = get_cpu();
477	idle_time = get_cpu_idle_time_us(cpu, NULL);
478	put_cpu();
479	if (idle_time != -1ULL) {
480		/* Idle micro accounting is supported. Use finer thresholds */
481		tuners->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
482		/*
483		 * In nohz/micro accounting case we set the minimum frequency
484		 * not depending on HZ, but fixed (very low). The deferred
485		 * timer might skip some samples if idle/sleeping as needed.
486		*/
487		dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
488	} else {
489		tuners->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
490
491		/* For correct statistics, we need 10 ticks for each measure */
492		dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
493			jiffies_to_usecs(10);
494	}
495
496	tuners->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
497	tuners->ignore_nice_load = 0;
498	tuners->powersave_bias = default_powersave_bias;
499	tuners->io_is_busy = should_io_be_busy();
500
501	dbs_data->tuners = tuners;
502	return 0;
503}
504
505static void od_exit(struct dbs_data *dbs_data, bool notify)
506{
507	kfree(dbs_data->tuners);
508}
509
510define_get_cpu_dbs_routines(od_cpu_dbs_info);
511
512static struct od_ops od_ops = {
513	.powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
514	.powersave_bias_target = generic_powersave_bias_target,
515	.freq_increase = dbs_freq_increase,
516};
517
518static struct common_dbs_data od_dbs_cdata = {
519	.governor = GOV_ONDEMAND,
520	.attr_group_gov_sys = &od_attr_group_gov_sys,
521	.attr_group_gov_pol = &od_attr_group_gov_pol,
522	.get_cpu_cdbs = get_cpu_cdbs,
523	.get_cpu_dbs_info_s = get_cpu_dbs_info_s,
524	.gov_dbs_timer = od_dbs_timer,
525	.gov_check_cpu = od_check_cpu,
526	.gov_ops = &od_ops,
527	.init = od_init,
528	.exit = od_exit,
529	.mutex = __MUTEX_INITIALIZER(od_dbs_cdata.mutex),
530};
531
532static void od_set_powersave_bias(unsigned int powersave_bias)
533{
534	struct cpufreq_policy *policy;
535	struct dbs_data *dbs_data;
536	struct od_dbs_tuners *od_tuners;
537	unsigned int cpu;
538	cpumask_t done;
539
540	default_powersave_bias = powersave_bias;
541	cpumask_clear(&done);
542
543	get_online_cpus();
544	for_each_online_cpu(cpu) {
545		struct cpu_common_dbs_info *shared;
546
547		if (cpumask_test_cpu(cpu, &done))
548			continue;
549
550		shared = per_cpu(od_cpu_dbs_info, cpu).cdbs.shared;
551		if (!shared)
552			continue;
553
554		policy = shared->policy;
555		cpumask_or(&done, &done, policy->cpus);
556
557		if (policy->governor != &cpufreq_gov_ondemand)
558			continue;
559
560		dbs_data = policy->governor_data;
561		od_tuners = dbs_data->tuners;
562		od_tuners->powersave_bias = default_powersave_bias;
563	}
564	put_online_cpus();
565}
566
567void od_register_powersave_bias_handler(unsigned int (*f)
568		(struct cpufreq_policy *, unsigned int, unsigned int),
569		unsigned int powersave_bias)
570{
571	od_ops.powersave_bias_target = f;
572	od_set_powersave_bias(powersave_bias);
573}
574EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
575
576void od_unregister_powersave_bias_handler(void)
577{
578	od_ops.powersave_bias_target = generic_powersave_bias_target;
579	od_set_powersave_bias(0);
580}
581EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
582
583static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
584		unsigned int event)
585{
586	return cpufreq_governor_dbs(policy, &od_dbs_cdata, event);
587}
588
589#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
590static
591#endif
592struct cpufreq_governor cpufreq_gov_ondemand = {
593	.name			= "ondemand",
594	.governor		= od_cpufreq_governor_dbs,
595	.max_transition_latency	= TRANSITION_LATENCY_LIMIT,
596	.owner			= THIS_MODULE,
597};
598
599static int __init cpufreq_gov_dbs_init(void)
600{
601	return cpufreq_register_governor(&cpufreq_gov_ondemand);
602}
603
604static void __exit cpufreq_gov_dbs_exit(void)
605{
606	cpufreq_unregister_governor(&cpufreq_gov_ondemand);
607}
608
609MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
610MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
611MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
612	"Low Latency Frequency Transition capable processors");
613MODULE_LICENSE("GPL");
614
615#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
616fs_initcall(cpufreq_gov_dbs_init);
617#else
618module_init(cpufreq_gov_dbs_init);
619#endif
620module_exit(cpufreq_gov_dbs_exit);
621