Linux Audio

Check our new training course

Loading...
v3.1
  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#include <linux/kernel.h>
 14#include <linux/module.h>
 15#include <linux/init.h>
 16#include <linux/cpufreq.h>
 17#include <linux/cpu.h>
 18#include <linux/jiffies.h>
 19#include <linux/kernel_stat.h>
 20#include <linux/mutex.h>
 21#include <linux/hrtimer.h>
 22#include <linux/tick.h>
 23#include <linux/ktime.h>
 24#include <linux/sched.h>
 25
 26/*
 27 * dbs is used in this file as a shortform for demandbased switching
 28 * It helps to keep variable names smaller, simpler
 29 */
 30
 31#define DEF_FREQUENCY_DOWN_DIFFERENTIAL		(10)
 32#define DEF_FREQUENCY_UP_THRESHOLD		(80)
 33#define DEF_SAMPLING_DOWN_FACTOR		(1)
 34#define MAX_SAMPLING_DOWN_FACTOR		(100000)
 35#define MICRO_FREQUENCY_DOWN_DIFFERENTIAL	(3)
 36#define MICRO_FREQUENCY_UP_THRESHOLD		(95)
 37#define MICRO_FREQUENCY_MIN_SAMPLE_RATE		(10000)
 38#define MIN_FREQUENCY_UP_THRESHOLD		(11)
 39#define MAX_FREQUENCY_UP_THRESHOLD		(100)
 40
 41/*
 42 * The polling frequency of this governor depends on the capability of
 43 * the processor. Default polling frequency is 1000 times the transition
 44 * latency of the processor. The governor will work on any processor with
 45 * transition latency <= 10mS, using appropriate sampling
 46 * rate.
 47 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
 48 * this governor will not work.
 49 * All times here are in uS.
 50 */
 51#define MIN_SAMPLING_RATE_RATIO			(2)
 52
 53static unsigned int min_sampling_rate;
 54
 55#define LATENCY_MULTIPLIER			(1000)
 56#define MIN_LATENCY_MULTIPLIER			(100)
 57#define TRANSITION_LATENCY_LIMIT		(10 * 1000 * 1000)
 58
 59static void do_dbs_timer(struct work_struct *work);
 60static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 61				unsigned int event);
 62
 63#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 64static
 65#endif
 66struct cpufreq_governor cpufreq_gov_ondemand = {
 67       .name                   = "ondemand",
 68       .governor               = cpufreq_governor_dbs,
 69       .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 70       .owner                  = THIS_MODULE,
 71};
 72
 73/* Sampling types */
 74enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
 75
 76struct cpu_dbs_info_s {
 77	cputime64_t prev_cpu_idle;
 78	cputime64_t prev_cpu_iowait;
 79	cputime64_t prev_cpu_wall;
 80	cputime64_t prev_cpu_nice;
 81	struct cpufreq_policy *cur_policy;
 82	struct delayed_work work;
 83	struct cpufreq_frequency_table *freq_table;
 84	unsigned int freq_lo;
 85	unsigned int freq_lo_jiffies;
 86	unsigned int freq_hi_jiffies;
 87	unsigned int rate_mult;
 88	int cpu;
 89	unsigned int sample_type:1;
 90	/*
 91	 * percpu mutex that serializes governor limit change with
 92	 * do_dbs_timer invocation. We do not want do_dbs_timer to run
 93	 * when user is changing the governor or limits.
 94	 */
 95	struct mutex timer_mutex;
 96};
 97static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
 98
 99static unsigned int dbs_enable;	/* number of CPUs using this policy */
100
101/*
102 * dbs_mutex protects dbs_enable in governor start/stop.
103 */
104static DEFINE_MUTEX(dbs_mutex);
105
106static struct dbs_tuners {
107	unsigned int sampling_rate;
108	unsigned int up_threshold;
109	unsigned int down_differential;
110	unsigned int ignore_nice;
111	unsigned int sampling_down_factor;
112	unsigned int powersave_bias;
113	unsigned int io_is_busy;
114} dbs_tuners_ins = {
115	.up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
116	.sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
117	.down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
118	.ignore_nice = 0,
119	.powersave_bias = 0,
120};
121
122static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
123							cputime64_t *wall)
124{
125	cputime64_t idle_time;
126	cputime64_t cur_wall_time;
127	cputime64_t busy_time;
128
129	cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
130	busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
131			kstat_cpu(cpu).cpustat.system);
132
133	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
134	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
135	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
136	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 
 
137
138	idle_time = cputime64_sub(cur_wall_time, busy_time);
139	if (wall)
140		*wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
141
142	return (cputime64_t)jiffies_to_usecs(idle_time);
143}
144
145static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
146{
147	u64 idle_time = get_cpu_idle_time_us(cpu, wall);
148
149	if (idle_time == -1ULL)
150		return get_cpu_idle_time_jiffy(cpu, wall);
 
 
151
152	return idle_time;
153}
154
155static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
156{
157	u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
158
159	if (iowait_time == -1ULL)
160		return 0;
161
162	return iowait_time;
163}
164
165/*
166 * Find right freq to be set now with powersave_bias on.
167 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
168 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
169 */
170static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
171					  unsigned int freq_next,
172					  unsigned int relation)
173{
174	unsigned int freq_req, freq_reduc, freq_avg;
175	unsigned int freq_hi, freq_lo;
176	unsigned int index = 0;
177	unsigned int jiffies_total, jiffies_hi, jiffies_lo;
178	struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
179						   policy->cpu);
180
181	if (!dbs_info->freq_table) {
182		dbs_info->freq_lo = 0;
183		dbs_info->freq_lo_jiffies = 0;
184		return freq_next;
185	}
186
187	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
188			relation, &index);
189	freq_req = dbs_info->freq_table[index].frequency;
190	freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
191	freq_avg = freq_req - freq_reduc;
192
193	/* Find freq bounds for freq_avg in freq_table */
194	index = 0;
195	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
196			CPUFREQ_RELATION_H, &index);
197	freq_lo = dbs_info->freq_table[index].frequency;
198	index = 0;
199	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
200			CPUFREQ_RELATION_L, &index);
201	freq_hi = dbs_info->freq_table[index].frequency;
202
203	/* Find out how long we have to be in hi and lo freqs */
204	if (freq_hi == freq_lo) {
205		dbs_info->freq_lo = 0;
206		dbs_info->freq_lo_jiffies = 0;
207		return freq_lo;
208	}
209	jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
210	jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
211	jiffies_hi += ((freq_hi - freq_lo) / 2);
212	jiffies_hi /= (freq_hi - freq_lo);
213	jiffies_lo = jiffies_total - jiffies_hi;
214	dbs_info->freq_lo = freq_lo;
215	dbs_info->freq_lo_jiffies = jiffies_lo;
216	dbs_info->freq_hi_jiffies = jiffies_hi;
217	return freq_hi;
218}
219
220static void ondemand_powersave_bias_init_cpu(int cpu)
221{
222	struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
223	dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
224	dbs_info->freq_lo = 0;
225}
226
227static void ondemand_powersave_bias_init(void)
228{
229	int i;
230	for_each_online_cpu(i) {
231		ondemand_powersave_bias_init_cpu(i);
232	}
233}
234
235/************************** sysfs interface ************************/
236
237static ssize_t show_sampling_rate_min(struct kobject *kobj,
238				      struct attribute *attr, char *buf)
239{
240	return sprintf(buf, "%u\n", min_sampling_rate);
241}
242
243define_one_global_ro(sampling_rate_min);
244
245/* cpufreq_ondemand Governor Tunables */
246#define show_one(file_name, object)					\
247static ssize_t show_##file_name						\
248(struct kobject *kobj, struct attribute *attr, char *buf)              \
249{									\
250	return sprintf(buf, "%u\n", dbs_tuners_ins.object);		\
251}
252show_one(sampling_rate, sampling_rate);
253show_one(io_is_busy, io_is_busy);
254show_one(up_threshold, up_threshold);
255show_one(sampling_down_factor, sampling_down_factor);
256show_one(ignore_nice_load, ignore_nice);
257show_one(powersave_bias, powersave_bias);
258
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
259static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
260				   const char *buf, size_t count)
261{
262	unsigned int input;
263	int ret;
264	ret = sscanf(buf, "%u", &input);
265	if (ret != 1)
266		return -EINVAL;
267	dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
268	return count;
269}
270
271static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
272				   const char *buf, size_t count)
273{
274	unsigned int input;
275	int ret;
276
277	ret = sscanf(buf, "%u", &input);
278	if (ret != 1)
279		return -EINVAL;
280	dbs_tuners_ins.io_is_busy = !!input;
281	return count;
282}
283
284static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
285				  const char *buf, size_t count)
286{
287	unsigned int input;
288	int ret;
289	ret = sscanf(buf, "%u", &input);
290
291	if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
292			input < MIN_FREQUENCY_UP_THRESHOLD) {
293		return -EINVAL;
294	}
295	dbs_tuners_ins.up_threshold = input;
296	return count;
297}
298
299static ssize_t store_sampling_down_factor(struct kobject *a,
300			struct attribute *b, const char *buf, size_t count)
301{
302	unsigned int input, j;
303	int ret;
304	ret = sscanf(buf, "%u", &input);
305
306	if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
307		return -EINVAL;
308	dbs_tuners_ins.sampling_down_factor = input;
309
310	/* Reset down sampling multiplier in case it was active */
311	for_each_online_cpu(j) {
312		struct cpu_dbs_info_s *dbs_info;
313		dbs_info = &per_cpu(od_cpu_dbs_info, j);
314		dbs_info->rate_mult = 1;
315	}
316	return count;
317}
318
319static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
320				      const char *buf, size_t count)
321{
322	unsigned int input;
323	int ret;
324
325	unsigned int j;
326
327	ret = sscanf(buf, "%u", &input);
328	if (ret != 1)
329		return -EINVAL;
330
331	if (input > 1)
332		input = 1;
333
334	if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
335		return count;
336	}
337	dbs_tuners_ins.ignore_nice = input;
338
339	/* we need to re-evaluate prev_cpu_idle */
340	for_each_online_cpu(j) {
341		struct cpu_dbs_info_s *dbs_info;
342		dbs_info = &per_cpu(od_cpu_dbs_info, j);
343		dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
344						&dbs_info->prev_cpu_wall);
345		if (dbs_tuners_ins.ignore_nice)
346			dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
347
348	}
349	return count;
350}
351
352static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
353				    const char *buf, size_t count)
354{
355	unsigned int input;
356	int ret;
357	ret = sscanf(buf, "%u", &input);
358
359	if (ret != 1)
360		return -EINVAL;
361
362	if (input > 1000)
363		input = 1000;
364
365	dbs_tuners_ins.powersave_bias = input;
366	ondemand_powersave_bias_init();
367	return count;
368}
369
370define_one_global_rw(sampling_rate);
371define_one_global_rw(io_is_busy);
372define_one_global_rw(up_threshold);
373define_one_global_rw(sampling_down_factor);
374define_one_global_rw(ignore_nice_load);
375define_one_global_rw(powersave_bias);
376
377static struct attribute *dbs_attributes[] = {
378	&sampling_rate_min.attr,
379	&sampling_rate.attr,
380	&up_threshold.attr,
381	&sampling_down_factor.attr,
382	&ignore_nice_load.attr,
383	&powersave_bias.attr,
384	&io_is_busy.attr,
385	NULL
386};
387
388static struct attribute_group dbs_attr_group = {
389	.attrs = dbs_attributes,
390	.name = "ondemand",
391};
392
393/************************** sysfs end ************************/
394
395static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
396{
397	if (dbs_tuners_ins.powersave_bias)
398		freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
399	else if (p->cur == p->max)
400		return;
401
402	__cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
403			CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
404}
405
406static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
407{
408	unsigned int max_load_freq;
409
410	struct cpufreq_policy *policy;
411	unsigned int j;
412
413	this_dbs_info->freq_lo = 0;
414	policy = this_dbs_info->cur_policy;
415
416	/*
417	 * Every sampling_rate, we check, if current idle time is less
418	 * than 20% (default), then we try to increase frequency
419	 * Every sampling_rate, we look for a the lowest
420	 * frequency which can sustain the load while keeping idle time over
421	 * 30%. If such a frequency exist, we try to decrease to this frequency.
422	 *
423	 * Any frequency increase takes it to the maximum frequency.
424	 * Frequency reduction happens at minimum steps of
425	 * 5% (default) of current frequency
426	 */
427
428	/* Get Absolute Load - in terms of freq */
429	max_load_freq = 0;
430
431	for_each_cpu(j, policy->cpus) {
432		struct cpu_dbs_info_s *j_dbs_info;
433		cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
434		unsigned int idle_time, wall_time, iowait_time;
435		unsigned int load, load_freq;
436		int freq_avg;
437
438		j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
439
440		cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
441		cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
442
443		wall_time = (unsigned int) cputime64_sub(cur_wall_time,
444				j_dbs_info->prev_cpu_wall);
445		j_dbs_info->prev_cpu_wall = cur_wall_time;
446
447		idle_time = (unsigned int) cputime64_sub(cur_idle_time,
448				j_dbs_info->prev_cpu_idle);
449		j_dbs_info->prev_cpu_idle = cur_idle_time;
450
451		iowait_time = (unsigned int) cputime64_sub(cur_iowait_time,
452				j_dbs_info->prev_cpu_iowait);
453		j_dbs_info->prev_cpu_iowait = cur_iowait_time;
454
455		if (dbs_tuners_ins.ignore_nice) {
456			cputime64_t cur_nice;
457			unsigned long cur_nice_jiffies;
458
459			cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
460					 j_dbs_info->prev_cpu_nice);
461			/*
462			 * Assumption: nice time between sampling periods will
463			 * be less than 2^32 jiffies for 32 bit sys
464			 */
465			cur_nice_jiffies = (unsigned long)
466					cputime64_to_jiffies64(cur_nice);
467
468			j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
469			idle_time += jiffies_to_usecs(cur_nice_jiffies);
470		}
471
472		/*
473		 * For the purpose of ondemand, waiting for disk IO is an
474		 * indication that you're performance critical, and not that
475		 * the system is actually idle. So subtract the iowait time
476		 * from the cpu idle time.
477		 */
478
479		if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
480			idle_time -= iowait_time;
481
482		if (unlikely(!wall_time || wall_time < idle_time))
483			continue;
484
485		load = 100 * (wall_time - idle_time) / wall_time;
486
487		freq_avg = __cpufreq_driver_getavg(policy, j);
488		if (freq_avg <= 0)
489			freq_avg = policy->cur;
490
491		load_freq = load * freq_avg;
492		if (load_freq > max_load_freq)
493			max_load_freq = load_freq;
494	}
495
496	/* Check for frequency increase */
497	if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
498		/* If switching to max speed, apply sampling_down_factor */
499		if (policy->cur < policy->max)
500			this_dbs_info->rate_mult =
501				dbs_tuners_ins.sampling_down_factor;
502		dbs_freq_increase(policy, policy->max);
503		return;
504	}
505
506	/* Check for frequency decrease */
507	/* if we cannot reduce the frequency anymore, break out early */
508	if (policy->cur == policy->min)
509		return;
510
511	/*
512	 * The optimal frequency is the frequency that is the lowest that
513	 * can support the current CPU usage without triggering the up
514	 * policy. To be safe, we focus 10 points under the threshold.
515	 */
516	if (max_load_freq <
517	    (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
518	     policy->cur) {
519		unsigned int freq_next;
520		freq_next = max_load_freq /
521				(dbs_tuners_ins.up_threshold -
522				 dbs_tuners_ins.down_differential);
523
524		/* No longer fully busy, reset rate_mult */
525		this_dbs_info->rate_mult = 1;
526
527		if (freq_next < policy->min)
528			freq_next = policy->min;
529
530		if (!dbs_tuners_ins.powersave_bias) {
531			__cpufreq_driver_target(policy, freq_next,
532					CPUFREQ_RELATION_L);
533		} else {
534			int freq = powersave_bias_target(policy, freq_next,
535					CPUFREQ_RELATION_L);
536			__cpufreq_driver_target(policy, freq,
537				CPUFREQ_RELATION_L);
538		}
539	}
540}
541
542static void do_dbs_timer(struct work_struct *work)
543{
544	struct cpu_dbs_info_s *dbs_info =
545		container_of(work, struct cpu_dbs_info_s, work.work);
546	unsigned int cpu = dbs_info->cpu;
547	int sample_type = dbs_info->sample_type;
548
549	int delay;
550
551	mutex_lock(&dbs_info->timer_mutex);
552
553	/* Common NORMAL_SAMPLE setup */
554	dbs_info->sample_type = DBS_NORMAL_SAMPLE;
555	if (!dbs_tuners_ins.powersave_bias ||
556	    sample_type == DBS_NORMAL_SAMPLE) {
557		dbs_check_cpu(dbs_info);
558		if (dbs_info->freq_lo) {
559			/* Setup timer for SUB_SAMPLE */
560			dbs_info->sample_type = DBS_SUB_SAMPLE;
561			delay = dbs_info->freq_hi_jiffies;
562		} else {
563			/* We want all CPUs to do sampling nearly on
564			 * same jiffy
565			 */
566			delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
567				* dbs_info->rate_mult);
568
569			if (num_online_cpus() > 1)
570				delay -= jiffies % delay;
571		}
572	} else {
573		__cpufreq_driver_target(dbs_info->cur_policy,
574			dbs_info->freq_lo, CPUFREQ_RELATION_H);
575		delay = dbs_info->freq_lo_jiffies;
576	}
577	schedule_delayed_work_on(cpu, &dbs_info->work, delay);
578	mutex_unlock(&dbs_info->timer_mutex);
579}
580
581static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
582{
583	/* We want all CPUs to do sampling nearly on same jiffy */
584	int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
585
586	if (num_online_cpus() > 1)
587		delay -= jiffies % delay;
588
589	dbs_info->sample_type = DBS_NORMAL_SAMPLE;
590	INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
591	schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
592}
593
594static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
595{
596	cancel_delayed_work_sync(&dbs_info->work);
597}
598
599/*
600 * Not all CPUs want IO time to be accounted as busy; this dependson how
601 * efficient idling at a higher frequency/voltage is.
602 * Pavel Machek says this is not so for various generations of AMD and old
603 * Intel systems.
604 * Mike Chan (androidlcom) calis this is also not true for ARM.
605 * Because of this, whitelist specific known (series) of CPUs by default, and
606 * leave all others up to the user.
607 */
608static int should_io_be_busy(void)
609{
610#if defined(CONFIG_X86)
611	/*
612	 * For Intel, Core 2 (model 15) andl later have an efficient idle.
613	 */
614	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
615	    boot_cpu_data.x86 == 6 &&
616	    boot_cpu_data.x86_model >= 15)
617		return 1;
618#endif
619	return 0;
620}
621
622static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
623				   unsigned int event)
624{
625	unsigned int cpu = policy->cpu;
626	struct cpu_dbs_info_s *this_dbs_info;
627	unsigned int j;
628	int rc;
629
630	this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
631
632	switch (event) {
633	case CPUFREQ_GOV_START:
634		if ((!cpu_online(cpu)) || (!policy->cur))
635			return -EINVAL;
636
637		mutex_lock(&dbs_mutex);
638
639		dbs_enable++;
640		for_each_cpu(j, policy->cpus) {
641			struct cpu_dbs_info_s *j_dbs_info;
642			j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
643			j_dbs_info->cur_policy = policy;
644
645			j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
646						&j_dbs_info->prev_cpu_wall);
647			if (dbs_tuners_ins.ignore_nice) {
648				j_dbs_info->prev_cpu_nice =
649						kstat_cpu(j).cpustat.nice;
650			}
651		}
652		this_dbs_info->cpu = cpu;
653		this_dbs_info->rate_mult = 1;
654		ondemand_powersave_bias_init_cpu(cpu);
655		/*
656		 * Start the timerschedule work, when this governor
657		 * is used for first time
658		 */
659		if (dbs_enable == 1) {
660			unsigned int latency;
661
662			rc = sysfs_create_group(cpufreq_global_kobject,
663						&dbs_attr_group);
664			if (rc) {
665				mutex_unlock(&dbs_mutex);
666				return rc;
667			}
668
669			/* policy latency is in nS. Convert it to uS first */
670			latency = policy->cpuinfo.transition_latency / 1000;
671			if (latency == 0)
672				latency = 1;
673			/* Bring kernel and HW constraints together */
674			min_sampling_rate = max(min_sampling_rate,
675					MIN_LATENCY_MULTIPLIER * latency);
676			dbs_tuners_ins.sampling_rate =
677				max(min_sampling_rate,
678				    latency * LATENCY_MULTIPLIER);
679			dbs_tuners_ins.io_is_busy = should_io_be_busy();
680		}
681		mutex_unlock(&dbs_mutex);
682
683		mutex_init(&this_dbs_info->timer_mutex);
684		dbs_timer_init(this_dbs_info);
685		break;
686
687	case CPUFREQ_GOV_STOP:
688		dbs_timer_exit(this_dbs_info);
689
690		mutex_lock(&dbs_mutex);
691		mutex_destroy(&this_dbs_info->timer_mutex);
692		dbs_enable--;
693		mutex_unlock(&dbs_mutex);
694		if (!dbs_enable)
695			sysfs_remove_group(cpufreq_global_kobject,
696					   &dbs_attr_group);
697
698		break;
699
700	case CPUFREQ_GOV_LIMITS:
701		mutex_lock(&this_dbs_info->timer_mutex);
702		if (policy->max < this_dbs_info->cur_policy->cur)
703			__cpufreq_driver_target(this_dbs_info->cur_policy,
704				policy->max, CPUFREQ_RELATION_H);
705		else if (policy->min > this_dbs_info->cur_policy->cur)
706			__cpufreq_driver_target(this_dbs_info->cur_policy,
707				policy->min, CPUFREQ_RELATION_L);
708		mutex_unlock(&this_dbs_info->timer_mutex);
709		break;
710	}
711	return 0;
712}
713
714static int __init cpufreq_gov_dbs_init(void)
715{
716	cputime64_t wall;
717	u64 idle_time;
718	int cpu = get_cpu();
719
720	idle_time = get_cpu_idle_time_us(cpu, &wall);
721	put_cpu();
722	if (idle_time != -1ULL) {
723		/* Idle micro accounting is supported. Use finer thresholds */
724		dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
725		dbs_tuners_ins.down_differential =
726					MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
727		/*
728		 * In no_hz/micro accounting case we set the minimum frequency
729		 * not depending on HZ, but fixed (very low). The deferred
730		 * timer might skip some samples if idle/sleeping as needed.
731		*/
732		min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
733	} else {
734		/* For correct statistics, we need 10 ticks for each measure */
735		min_sampling_rate =
736			MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
737	}
738
739	return cpufreq_register_governor(&cpufreq_gov_ondemand);
740}
741
742static void __exit cpufreq_gov_dbs_exit(void)
743{
744	cpufreq_unregister_governor(&cpufreq_gov_ondemand);
745}
746
747
748MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
749MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
750MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
751	"Low Latency Frequency Transition capable processors");
752MODULE_LICENSE("GPL");
753
754#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
755fs_initcall(cpufreq_gov_dbs_init);
756#else
757module_init(cpufreq_gov_dbs_init);
758#endif
759module_exit(cpufreq_gov_dbs_exit);
v3.5.6
  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#include <linux/kernel.h>
 14#include <linux/module.h>
 15#include <linux/init.h>
 16#include <linux/cpufreq.h>
 17#include <linux/cpu.h>
 18#include <linux/jiffies.h>
 19#include <linux/kernel_stat.h>
 20#include <linux/mutex.h>
 21#include <linux/hrtimer.h>
 22#include <linux/tick.h>
 23#include <linux/ktime.h>
 24#include <linux/sched.h>
 25
 26/*
 27 * dbs is used in this file as a shortform for demandbased switching
 28 * It helps to keep variable names smaller, simpler
 29 */
 30
 31#define DEF_FREQUENCY_DOWN_DIFFERENTIAL		(10)
 32#define DEF_FREQUENCY_UP_THRESHOLD		(80)
 33#define DEF_SAMPLING_DOWN_FACTOR		(1)
 34#define MAX_SAMPLING_DOWN_FACTOR		(100000)
 35#define MICRO_FREQUENCY_DOWN_DIFFERENTIAL	(3)
 36#define MICRO_FREQUENCY_UP_THRESHOLD		(95)
 37#define MICRO_FREQUENCY_MIN_SAMPLE_RATE		(10000)
 38#define MIN_FREQUENCY_UP_THRESHOLD		(11)
 39#define MAX_FREQUENCY_UP_THRESHOLD		(100)
 40
 41/*
 42 * The polling frequency of this governor depends on the capability of
 43 * the processor. Default polling frequency is 1000 times the transition
 44 * latency of the processor. The governor will work on any processor with
 45 * transition latency <= 10mS, using appropriate sampling
 46 * rate.
 47 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
 48 * this governor will not work.
 49 * All times here are in uS.
 50 */
 51#define MIN_SAMPLING_RATE_RATIO			(2)
 52
 53static unsigned int min_sampling_rate;
 54
 55#define LATENCY_MULTIPLIER			(1000)
 56#define MIN_LATENCY_MULTIPLIER			(100)
 57#define TRANSITION_LATENCY_LIMIT		(10 * 1000 * 1000)
 58
 59static void do_dbs_timer(struct work_struct *work);
 60static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 61				unsigned int event);
 62
 63#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 64static
 65#endif
 66struct cpufreq_governor cpufreq_gov_ondemand = {
 67       .name                   = "ondemand",
 68       .governor               = cpufreq_governor_dbs,
 69       .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 70       .owner                  = THIS_MODULE,
 71};
 72
 73/* Sampling types */
 74enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
 75
 76struct cpu_dbs_info_s {
 77	cputime64_t prev_cpu_idle;
 78	cputime64_t prev_cpu_iowait;
 79	cputime64_t prev_cpu_wall;
 80	cputime64_t prev_cpu_nice;
 81	struct cpufreq_policy *cur_policy;
 82	struct delayed_work work;
 83	struct cpufreq_frequency_table *freq_table;
 84	unsigned int freq_lo;
 85	unsigned int freq_lo_jiffies;
 86	unsigned int freq_hi_jiffies;
 87	unsigned int rate_mult;
 88	int cpu;
 89	unsigned int sample_type:1;
 90	/*
 91	 * percpu mutex that serializes governor limit change with
 92	 * do_dbs_timer invocation. We do not want do_dbs_timer to run
 93	 * when user is changing the governor or limits.
 94	 */
 95	struct mutex timer_mutex;
 96};
 97static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
 98
 99static unsigned int dbs_enable;	/* number of CPUs using this policy */
100
101/*
102 * dbs_mutex protects dbs_enable in governor start/stop.
103 */
104static DEFINE_MUTEX(dbs_mutex);
105
106static struct dbs_tuners {
107	unsigned int sampling_rate;
108	unsigned int up_threshold;
109	unsigned int down_differential;
110	unsigned int ignore_nice;
111	unsigned int sampling_down_factor;
112	unsigned int powersave_bias;
113	unsigned int io_is_busy;
114} dbs_tuners_ins = {
115	.up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
116	.sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
117	.down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
118	.ignore_nice = 0,
119	.powersave_bias = 0,
120};
121
122static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
 
123{
124	u64 idle_time;
125	u64 cur_wall_time;
126	u64 busy_time;
127
128	cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 
 
129
130	busy_time  = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
131	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
132	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
133	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
134	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
135	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
136
137	idle_time = cur_wall_time - busy_time;
138	if (wall)
139		*wall = jiffies_to_usecs(cur_wall_time);
140
141	return jiffies_to_usecs(idle_time);
142}
143
144static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
145{
146	u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
147
148	if (idle_time == -1ULL)
149		return get_cpu_idle_time_jiffy(cpu, wall);
150	else
151		idle_time += get_cpu_iowait_time_us(cpu, wall);
152
153	return idle_time;
154}
155
156static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
157{
158	u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
159
160	if (iowait_time == -1ULL)
161		return 0;
162
163	return iowait_time;
164}
165
166/*
167 * Find right freq to be set now with powersave_bias on.
168 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
169 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
170 */
171static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
172					  unsigned int freq_next,
173					  unsigned int relation)
174{
175	unsigned int freq_req, freq_reduc, freq_avg;
176	unsigned int freq_hi, freq_lo;
177	unsigned int index = 0;
178	unsigned int jiffies_total, jiffies_hi, jiffies_lo;
179	struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
180						   policy->cpu);
181
182	if (!dbs_info->freq_table) {
183		dbs_info->freq_lo = 0;
184		dbs_info->freq_lo_jiffies = 0;
185		return freq_next;
186	}
187
188	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
189			relation, &index);
190	freq_req = dbs_info->freq_table[index].frequency;
191	freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
192	freq_avg = freq_req - freq_reduc;
193
194	/* Find freq bounds for freq_avg in freq_table */
195	index = 0;
196	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
197			CPUFREQ_RELATION_H, &index);
198	freq_lo = dbs_info->freq_table[index].frequency;
199	index = 0;
200	cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
201			CPUFREQ_RELATION_L, &index);
202	freq_hi = dbs_info->freq_table[index].frequency;
203
204	/* Find out how long we have to be in hi and lo freqs */
205	if (freq_hi == freq_lo) {
206		dbs_info->freq_lo = 0;
207		dbs_info->freq_lo_jiffies = 0;
208		return freq_lo;
209	}
210	jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
211	jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
212	jiffies_hi += ((freq_hi - freq_lo) / 2);
213	jiffies_hi /= (freq_hi - freq_lo);
214	jiffies_lo = jiffies_total - jiffies_hi;
215	dbs_info->freq_lo = freq_lo;
216	dbs_info->freq_lo_jiffies = jiffies_lo;
217	dbs_info->freq_hi_jiffies = jiffies_hi;
218	return freq_hi;
219}
220
221static void ondemand_powersave_bias_init_cpu(int cpu)
222{
223	struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
224	dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
225	dbs_info->freq_lo = 0;
226}
227
228static void ondemand_powersave_bias_init(void)
229{
230	int i;
231	for_each_online_cpu(i) {
232		ondemand_powersave_bias_init_cpu(i);
233	}
234}
235
236/************************** sysfs interface ************************/
237
238static ssize_t show_sampling_rate_min(struct kobject *kobj,
239				      struct attribute *attr, char *buf)
240{
241	return sprintf(buf, "%u\n", min_sampling_rate);
242}
243
244define_one_global_ro(sampling_rate_min);
245
246/* cpufreq_ondemand Governor Tunables */
247#define show_one(file_name, object)					\
248static ssize_t show_##file_name						\
249(struct kobject *kobj, struct attribute *attr, char *buf)              \
250{									\
251	return sprintf(buf, "%u\n", dbs_tuners_ins.object);		\
252}
253show_one(sampling_rate, sampling_rate);
254show_one(io_is_busy, io_is_busy);
255show_one(up_threshold, up_threshold);
256show_one(sampling_down_factor, sampling_down_factor);
257show_one(ignore_nice_load, ignore_nice);
258show_one(powersave_bias, powersave_bias);
259
260/**
261 * update_sampling_rate - update sampling rate effective immediately if needed.
262 * @new_rate: new sampling rate
263 *
264 * If new rate is smaller than the old, simply updaing
265 * dbs_tuners_int.sampling_rate might not be appropriate. For example,
266 * if the original sampling_rate was 1 second and the requested new sampling
267 * rate is 10 ms because the user needs immediate reaction from ondemand
268 * governor, but not sure if higher frequency will be required or not,
269 * then, the governor may change the sampling rate too late; up to 1 second
270 * later. Thus, if we are reducing the sampling rate, we need to make the
271 * new value effective immediately.
272 */
273static void update_sampling_rate(unsigned int new_rate)
274{
275	int cpu;
276
277	dbs_tuners_ins.sampling_rate = new_rate
278				     = max(new_rate, min_sampling_rate);
279
280	for_each_online_cpu(cpu) {
281		struct cpufreq_policy *policy;
282		struct cpu_dbs_info_s *dbs_info;
283		unsigned long next_sampling, appointed_at;
284
285		policy = cpufreq_cpu_get(cpu);
286		if (!policy)
287			continue;
288		dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
289		cpufreq_cpu_put(policy);
290
291		mutex_lock(&dbs_info->timer_mutex);
292
293		if (!delayed_work_pending(&dbs_info->work)) {
294			mutex_unlock(&dbs_info->timer_mutex);
295			continue;
296		}
297
298		next_sampling  = jiffies + usecs_to_jiffies(new_rate);
299		appointed_at = dbs_info->work.timer.expires;
300
301
302		if (time_before(next_sampling, appointed_at)) {
303
304			mutex_unlock(&dbs_info->timer_mutex);
305			cancel_delayed_work_sync(&dbs_info->work);
306			mutex_lock(&dbs_info->timer_mutex);
307
308			schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work,
309						 usecs_to_jiffies(new_rate));
310
311		}
312		mutex_unlock(&dbs_info->timer_mutex);
313	}
314}
315
316static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
317				   const char *buf, size_t count)
318{
319	unsigned int input;
320	int ret;
321	ret = sscanf(buf, "%u", &input);
322	if (ret != 1)
323		return -EINVAL;
324	update_sampling_rate(input);
325	return count;
326}
327
328static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
329				   const char *buf, size_t count)
330{
331	unsigned int input;
332	int ret;
333
334	ret = sscanf(buf, "%u", &input);
335	if (ret != 1)
336		return -EINVAL;
337	dbs_tuners_ins.io_is_busy = !!input;
338	return count;
339}
340
341static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
342				  const char *buf, size_t count)
343{
344	unsigned int input;
345	int ret;
346	ret = sscanf(buf, "%u", &input);
347
348	if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
349			input < MIN_FREQUENCY_UP_THRESHOLD) {
350		return -EINVAL;
351	}
352	dbs_tuners_ins.up_threshold = input;
353	return count;
354}
355
356static ssize_t store_sampling_down_factor(struct kobject *a,
357			struct attribute *b, const char *buf, size_t count)
358{
359	unsigned int input, j;
360	int ret;
361	ret = sscanf(buf, "%u", &input);
362
363	if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
364		return -EINVAL;
365	dbs_tuners_ins.sampling_down_factor = input;
366
367	/* Reset down sampling multiplier in case it was active */
368	for_each_online_cpu(j) {
369		struct cpu_dbs_info_s *dbs_info;
370		dbs_info = &per_cpu(od_cpu_dbs_info, j);
371		dbs_info->rate_mult = 1;
372	}
373	return count;
374}
375
376static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
377				      const char *buf, size_t count)
378{
379	unsigned int input;
380	int ret;
381
382	unsigned int j;
383
384	ret = sscanf(buf, "%u", &input);
385	if (ret != 1)
386		return -EINVAL;
387
388	if (input > 1)
389		input = 1;
390
391	if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
392		return count;
393	}
394	dbs_tuners_ins.ignore_nice = input;
395
396	/* we need to re-evaluate prev_cpu_idle */
397	for_each_online_cpu(j) {
398		struct cpu_dbs_info_s *dbs_info;
399		dbs_info = &per_cpu(od_cpu_dbs_info, j);
400		dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
401						&dbs_info->prev_cpu_wall);
402		if (dbs_tuners_ins.ignore_nice)
403			dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
404
405	}
406	return count;
407}
408
409static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
410				    const char *buf, size_t count)
411{
412	unsigned int input;
413	int ret;
414	ret = sscanf(buf, "%u", &input);
415
416	if (ret != 1)
417		return -EINVAL;
418
419	if (input > 1000)
420		input = 1000;
421
422	dbs_tuners_ins.powersave_bias = input;
423	ondemand_powersave_bias_init();
424	return count;
425}
426
427define_one_global_rw(sampling_rate);
428define_one_global_rw(io_is_busy);
429define_one_global_rw(up_threshold);
430define_one_global_rw(sampling_down_factor);
431define_one_global_rw(ignore_nice_load);
432define_one_global_rw(powersave_bias);
433
434static struct attribute *dbs_attributes[] = {
435	&sampling_rate_min.attr,
436	&sampling_rate.attr,
437	&up_threshold.attr,
438	&sampling_down_factor.attr,
439	&ignore_nice_load.attr,
440	&powersave_bias.attr,
441	&io_is_busy.attr,
442	NULL
443};
444
445static struct attribute_group dbs_attr_group = {
446	.attrs = dbs_attributes,
447	.name = "ondemand",
448};
449
450/************************** sysfs end ************************/
451
452static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
453{
454	if (dbs_tuners_ins.powersave_bias)
455		freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
456	else if (p->cur == p->max)
457		return;
458
459	__cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
460			CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
461}
462
463static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
464{
465	unsigned int max_load_freq;
466
467	struct cpufreq_policy *policy;
468	unsigned int j;
469
470	this_dbs_info->freq_lo = 0;
471	policy = this_dbs_info->cur_policy;
472
473	/*
474	 * Every sampling_rate, we check, if current idle time is less
475	 * than 20% (default), then we try to increase frequency
476	 * Every sampling_rate, we look for a the lowest
477	 * frequency which can sustain the load while keeping idle time over
478	 * 30%. If such a frequency exist, we try to decrease to this frequency.
479	 *
480	 * Any frequency increase takes it to the maximum frequency.
481	 * Frequency reduction happens at minimum steps of
482	 * 5% (default) of current frequency
483	 */
484
485	/* Get Absolute Load - in terms of freq */
486	max_load_freq = 0;
487
488	for_each_cpu(j, policy->cpus) {
489		struct cpu_dbs_info_s *j_dbs_info;
490		cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
491		unsigned int idle_time, wall_time, iowait_time;
492		unsigned int load, load_freq;
493		int freq_avg;
494
495		j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
496
497		cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
498		cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
499
500		wall_time = (unsigned int)
501			(cur_wall_time - j_dbs_info->prev_cpu_wall);
502		j_dbs_info->prev_cpu_wall = cur_wall_time;
503
504		idle_time = (unsigned int)
505			(cur_idle_time - j_dbs_info->prev_cpu_idle);
506		j_dbs_info->prev_cpu_idle = cur_idle_time;
507
508		iowait_time = (unsigned int)
509			(cur_iowait_time - j_dbs_info->prev_cpu_iowait);
510		j_dbs_info->prev_cpu_iowait = cur_iowait_time;
511
512		if (dbs_tuners_ins.ignore_nice) {
513			u64 cur_nice;
514			unsigned long cur_nice_jiffies;
515
516			cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
517					 j_dbs_info->prev_cpu_nice;
518			/*
519			 * Assumption: nice time between sampling periods will
520			 * be less than 2^32 jiffies for 32 bit sys
521			 */
522			cur_nice_jiffies = (unsigned long)
523					cputime64_to_jiffies64(cur_nice);
524
525			j_dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
526			idle_time += jiffies_to_usecs(cur_nice_jiffies);
527		}
528
529		/*
530		 * For the purpose of ondemand, waiting for disk IO is an
531		 * indication that you're performance critical, and not that
532		 * the system is actually idle. So subtract the iowait time
533		 * from the cpu idle time.
534		 */
535
536		if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
537			idle_time -= iowait_time;
538
539		if (unlikely(!wall_time || wall_time < idle_time))
540			continue;
541
542		load = 100 * (wall_time - idle_time) / wall_time;
543
544		freq_avg = __cpufreq_driver_getavg(policy, j);
545		if (freq_avg <= 0)
546			freq_avg = policy->cur;
547
548		load_freq = load * freq_avg;
549		if (load_freq > max_load_freq)
550			max_load_freq = load_freq;
551	}
552
553	/* Check for frequency increase */
554	if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
555		/* If switching to max speed, apply sampling_down_factor */
556		if (policy->cur < policy->max)
557			this_dbs_info->rate_mult =
558				dbs_tuners_ins.sampling_down_factor;
559		dbs_freq_increase(policy, policy->max);
560		return;
561	}
562
563	/* Check for frequency decrease */
564	/* if we cannot reduce the frequency anymore, break out early */
565	if (policy->cur == policy->min)
566		return;
567
568	/*
569	 * The optimal frequency is the frequency that is the lowest that
570	 * can support the current CPU usage without triggering the up
571	 * policy. To be safe, we focus 10 points under the threshold.
572	 */
573	if (max_load_freq <
574	    (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
575	     policy->cur) {
576		unsigned int freq_next;
577		freq_next = max_load_freq /
578				(dbs_tuners_ins.up_threshold -
579				 dbs_tuners_ins.down_differential);
580
581		/* No longer fully busy, reset rate_mult */
582		this_dbs_info->rate_mult = 1;
583
584		if (freq_next < policy->min)
585			freq_next = policy->min;
586
587		if (!dbs_tuners_ins.powersave_bias) {
588			__cpufreq_driver_target(policy, freq_next,
589					CPUFREQ_RELATION_L);
590		} else {
591			int freq = powersave_bias_target(policy, freq_next,
592					CPUFREQ_RELATION_L);
593			__cpufreq_driver_target(policy, freq,
594				CPUFREQ_RELATION_L);
595		}
596	}
597}
598
599static void do_dbs_timer(struct work_struct *work)
600{
601	struct cpu_dbs_info_s *dbs_info =
602		container_of(work, struct cpu_dbs_info_s, work.work);
603	unsigned int cpu = dbs_info->cpu;
604	int sample_type = dbs_info->sample_type;
605
606	int delay;
607
608	mutex_lock(&dbs_info->timer_mutex);
609
610	/* Common NORMAL_SAMPLE setup */
611	dbs_info->sample_type = DBS_NORMAL_SAMPLE;
612	if (!dbs_tuners_ins.powersave_bias ||
613	    sample_type == DBS_NORMAL_SAMPLE) {
614		dbs_check_cpu(dbs_info);
615		if (dbs_info->freq_lo) {
616			/* Setup timer for SUB_SAMPLE */
617			dbs_info->sample_type = DBS_SUB_SAMPLE;
618			delay = dbs_info->freq_hi_jiffies;
619		} else {
620			/* We want all CPUs to do sampling nearly on
621			 * same jiffy
622			 */
623			delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
624				* dbs_info->rate_mult);
625
626			if (num_online_cpus() > 1)
627				delay -= jiffies % delay;
628		}
629	} else {
630		__cpufreq_driver_target(dbs_info->cur_policy,
631			dbs_info->freq_lo, CPUFREQ_RELATION_H);
632		delay = dbs_info->freq_lo_jiffies;
633	}
634	schedule_delayed_work_on(cpu, &dbs_info->work, delay);
635	mutex_unlock(&dbs_info->timer_mutex);
636}
637
638static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
639{
640	/* We want all CPUs to do sampling nearly on same jiffy */
641	int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
642
643	if (num_online_cpus() > 1)
644		delay -= jiffies % delay;
645
646	dbs_info->sample_type = DBS_NORMAL_SAMPLE;
647	INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
648	schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
649}
650
651static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
652{
653	cancel_delayed_work_sync(&dbs_info->work);
654}
655
656/*
657 * Not all CPUs want IO time to be accounted as busy; this dependson how
658 * efficient idling at a higher frequency/voltage is.
659 * Pavel Machek says this is not so for various generations of AMD and old
660 * Intel systems.
661 * Mike Chan (androidlcom) calis this is also not true for ARM.
662 * Because of this, whitelist specific known (series) of CPUs by default, and
663 * leave all others up to the user.
664 */
665static int should_io_be_busy(void)
666{
667#if defined(CONFIG_X86)
668	/*
669	 * For Intel, Core 2 (model 15) andl later have an efficient idle.
670	 */
671	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
672	    boot_cpu_data.x86 == 6 &&
673	    boot_cpu_data.x86_model >= 15)
674		return 1;
675#endif
676	return 0;
677}
678
679static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
680				   unsigned int event)
681{
682	unsigned int cpu = policy->cpu;
683	struct cpu_dbs_info_s *this_dbs_info;
684	unsigned int j;
685	int rc;
686
687	this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
688
689	switch (event) {
690	case CPUFREQ_GOV_START:
691		if ((!cpu_online(cpu)) || (!policy->cur))
692			return -EINVAL;
693
694		mutex_lock(&dbs_mutex);
695
696		dbs_enable++;
697		for_each_cpu(j, policy->cpus) {
698			struct cpu_dbs_info_s *j_dbs_info;
699			j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
700			j_dbs_info->cur_policy = policy;
701
702			j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
703						&j_dbs_info->prev_cpu_wall);
704			if (dbs_tuners_ins.ignore_nice)
705				j_dbs_info->prev_cpu_nice =
706						kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 
707		}
708		this_dbs_info->cpu = cpu;
709		this_dbs_info->rate_mult = 1;
710		ondemand_powersave_bias_init_cpu(cpu);
711		/*
712		 * Start the timerschedule work, when this governor
713		 * is used for first time
714		 */
715		if (dbs_enable == 1) {
716			unsigned int latency;
717
718			rc = sysfs_create_group(cpufreq_global_kobject,
719						&dbs_attr_group);
720			if (rc) {
721				mutex_unlock(&dbs_mutex);
722				return rc;
723			}
724
725			/* policy latency is in nS. Convert it to uS first */
726			latency = policy->cpuinfo.transition_latency / 1000;
727			if (latency == 0)
728				latency = 1;
729			/* Bring kernel and HW constraints together */
730			min_sampling_rate = max(min_sampling_rate,
731					MIN_LATENCY_MULTIPLIER * latency);
732			dbs_tuners_ins.sampling_rate =
733				max(min_sampling_rate,
734				    latency * LATENCY_MULTIPLIER);
735			dbs_tuners_ins.io_is_busy = should_io_be_busy();
736		}
737		mutex_unlock(&dbs_mutex);
738
739		mutex_init(&this_dbs_info->timer_mutex);
740		dbs_timer_init(this_dbs_info);
741		break;
742
743	case CPUFREQ_GOV_STOP:
744		dbs_timer_exit(this_dbs_info);
745
746		mutex_lock(&dbs_mutex);
747		mutex_destroy(&this_dbs_info->timer_mutex);
748		dbs_enable--;
749		mutex_unlock(&dbs_mutex);
750		if (!dbs_enable)
751			sysfs_remove_group(cpufreq_global_kobject,
752					   &dbs_attr_group);
753
754		break;
755
756	case CPUFREQ_GOV_LIMITS:
757		mutex_lock(&this_dbs_info->timer_mutex);
758		if (policy->max < this_dbs_info->cur_policy->cur)
759			__cpufreq_driver_target(this_dbs_info->cur_policy,
760				policy->max, CPUFREQ_RELATION_H);
761		else if (policy->min > this_dbs_info->cur_policy->cur)
762			__cpufreq_driver_target(this_dbs_info->cur_policy,
763				policy->min, CPUFREQ_RELATION_L);
764		mutex_unlock(&this_dbs_info->timer_mutex);
765		break;
766	}
767	return 0;
768}
769
770static int __init cpufreq_gov_dbs_init(void)
771{
 
772	u64 idle_time;
773	int cpu = get_cpu();
774
775	idle_time = get_cpu_idle_time_us(cpu, NULL);
776	put_cpu();
777	if (idle_time != -1ULL) {
778		/* Idle micro accounting is supported. Use finer thresholds */
779		dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
780		dbs_tuners_ins.down_differential =
781					MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
782		/*
783		 * In nohz/micro accounting case we set the minimum frequency
784		 * not depending on HZ, but fixed (very low). The deferred
785		 * timer might skip some samples if idle/sleeping as needed.
786		*/
787		min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
788	} else {
789		/* For correct statistics, we need 10 ticks for each measure */
790		min_sampling_rate =
791			MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
792	}
793
794	return cpufreq_register_governor(&cpufreq_gov_ondemand);
795}
796
797static void __exit cpufreq_gov_dbs_exit(void)
798{
799	cpufreq_unregister_governor(&cpufreq_gov_ondemand);
800}
801
802
803MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
804MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
805MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
806	"Low Latency Frequency Transition capable processors");
807MODULE_LICENSE("GPL");
808
809#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
810fs_initcall(cpufreq_gov_dbs_init);
811#else
812module_init(cpufreq_gov_dbs_init);
813#endif
814module_exit(cpufreq_gov_dbs_exit);