Loading...
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);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * drivers/cpufreq/cpufreq_ondemand.c
4 *
5 * Copyright (C) 2001 Russell King
6 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
7 * Jun Nakajima <jun.nakajima@intel.com>
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/cpu.h>
13#include <linux/percpu-defs.h>
14#include <linux/slab.h>
15#include <linux/tick.h>
16#include <linux/sched/cpufreq.h>
17
18#include "cpufreq_ondemand.h"
19
20/* On-demand governor macros */
21#define DEF_FREQUENCY_UP_THRESHOLD (80)
22#define DEF_SAMPLING_DOWN_FACTOR (1)
23#define MAX_SAMPLING_DOWN_FACTOR (100000)
24#define MICRO_FREQUENCY_UP_THRESHOLD (95)
25#define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
26#define MIN_FREQUENCY_UP_THRESHOLD (1)
27#define MAX_FREQUENCY_UP_THRESHOLD (100)
28
29static struct od_ops od_ops;
30
31static unsigned int default_powersave_bias;
32
33/*
34 * Not all CPUs want IO time to be accounted as busy; this depends on how
35 * efficient idling at a higher frequency/voltage is.
36 * Pavel Machek says this is not so for various generations of AMD and old
37 * Intel systems.
38 * Mike Chan (android.com) claims this is also not true for ARM.
39 * Because of this, whitelist specific known (series) of CPUs by default, and
40 * leave all others up to the user.
41 */
42static int should_io_be_busy(void)
43{
44#if defined(CONFIG_X86)
45 /*
46 * For Intel, Core 2 (model 15) and later have an efficient idle.
47 */
48 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
49 boot_cpu_data.x86 == 6 &&
50 boot_cpu_data.x86_model >= 15)
51 return 1;
52#endif
53 return 0;
54}
55
56/*
57 * Find right freq to be set now with powersave_bias on.
58 * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
59 * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
60 */
61static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
62 unsigned int freq_next, unsigned int relation)
63{
64 unsigned int freq_req, freq_reduc, freq_avg;
65 unsigned int freq_hi, freq_lo;
66 unsigned int index;
67 unsigned int delay_hi_us;
68 struct policy_dbs_info *policy_dbs = policy->governor_data;
69 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
70 struct dbs_data *dbs_data = policy_dbs->dbs_data;
71 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
72 struct cpufreq_frequency_table *freq_table = policy->freq_table;
73
74 if (!freq_table) {
75 dbs_info->freq_lo = 0;
76 dbs_info->freq_lo_delay_us = 0;
77 return freq_next;
78 }
79
80 index = cpufreq_frequency_table_target(policy, freq_next, relation);
81 freq_req = freq_table[index].frequency;
82 freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
83 freq_avg = freq_req - freq_reduc;
84
85 /* Find freq bounds for freq_avg in freq_table */
86 index = cpufreq_table_find_index_h(policy, freq_avg);
87 freq_lo = freq_table[index].frequency;
88 index = cpufreq_table_find_index_l(policy, freq_avg);
89 freq_hi = freq_table[index].frequency;
90
91 /* Find out how long we have to be in hi and lo freqs */
92 if (freq_hi == freq_lo) {
93 dbs_info->freq_lo = 0;
94 dbs_info->freq_lo_delay_us = 0;
95 return freq_lo;
96 }
97 delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
98 delay_hi_us += (freq_hi - freq_lo) / 2;
99 delay_hi_us /= freq_hi - freq_lo;
100 dbs_info->freq_hi_delay_us = delay_hi_us;
101 dbs_info->freq_lo = freq_lo;
102 dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
103 return freq_hi;
104}
105
106static void ondemand_powersave_bias_init(struct cpufreq_policy *policy)
107{
108 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
109
110 dbs_info->freq_lo = 0;
111}
112
113static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
114{
115 struct policy_dbs_info *policy_dbs = policy->governor_data;
116 struct dbs_data *dbs_data = policy_dbs->dbs_data;
117 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
118
119 if (od_tuners->powersave_bias)
120 freq = od_ops.powersave_bias_target(policy, freq,
121 CPUFREQ_RELATION_H);
122 else if (policy->cur == policy->max)
123 return;
124
125 __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
126 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
127}
128
129/*
130 * Every sampling_rate, we check, if current idle time is less than 20%
131 * (default), then we try to increase frequency. Else, we adjust the frequency
132 * proportional to load.
133 */
134static void od_update(struct cpufreq_policy *policy)
135{
136 struct policy_dbs_info *policy_dbs = policy->governor_data;
137 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
138 struct dbs_data *dbs_data = policy_dbs->dbs_data;
139 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
140 unsigned int load = dbs_update(policy);
141
142 dbs_info->freq_lo = 0;
143
144 /* Check for frequency increase */
145 if (load > dbs_data->up_threshold) {
146 /* If switching to max speed, apply sampling_down_factor */
147 if (policy->cur < policy->max)
148 policy_dbs->rate_mult = dbs_data->sampling_down_factor;
149 dbs_freq_increase(policy, policy->max);
150 } else {
151 /* Calculate the next frequency proportional to load */
152 unsigned int freq_next, min_f, max_f;
153
154 min_f = policy->cpuinfo.min_freq;
155 max_f = policy->cpuinfo.max_freq;
156 freq_next = min_f + load * (max_f - min_f) / 100;
157
158 /* No longer fully busy, reset rate_mult */
159 policy_dbs->rate_mult = 1;
160
161 if (od_tuners->powersave_bias)
162 freq_next = od_ops.powersave_bias_target(policy,
163 freq_next,
164 CPUFREQ_RELATION_L);
165
166 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
167 }
168}
169
170static unsigned int od_dbs_update(struct cpufreq_policy *policy)
171{
172 struct policy_dbs_info *policy_dbs = policy->governor_data;
173 struct dbs_data *dbs_data = policy_dbs->dbs_data;
174 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
175 int sample_type = dbs_info->sample_type;
176
177 /* Common NORMAL_SAMPLE setup */
178 dbs_info->sample_type = OD_NORMAL_SAMPLE;
179 /*
180 * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
181 * it then.
182 */
183 if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
184 __cpufreq_driver_target(policy, dbs_info->freq_lo,
185 CPUFREQ_RELATION_H);
186 return dbs_info->freq_lo_delay_us;
187 }
188
189 od_update(policy);
190
191 if (dbs_info->freq_lo) {
192 /* Setup SUB_SAMPLE */
193 dbs_info->sample_type = OD_SUB_SAMPLE;
194 return dbs_info->freq_hi_delay_us;
195 }
196
197 return dbs_data->sampling_rate * policy_dbs->rate_mult;
198}
199
200/************************** sysfs interface ************************/
201static struct dbs_governor od_dbs_gov;
202
203static ssize_t store_io_is_busy(struct gov_attr_set *attr_set, const char *buf,
204 size_t count)
205{
206 struct dbs_data *dbs_data = to_dbs_data(attr_set);
207 unsigned int input;
208 int ret;
209
210 ret = sscanf(buf, "%u", &input);
211 if (ret != 1)
212 return -EINVAL;
213 dbs_data->io_is_busy = !!input;
214
215 /* we need to re-evaluate prev_cpu_idle */
216 gov_update_cpu_data(dbs_data);
217
218 return count;
219}
220
221static ssize_t store_up_threshold(struct gov_attr_set *attr_set,
222 const char *buf, size_t count)
223{
224 struct dbs_data *dbs_data = to_dbs_data(attr_set);
225 unsigned int input;
226 int ret;
227 ret = sscanf(buf, "%u", &input);
228
229 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
230 input < MIN_FREQUENCY_UP_THRESHOLD) {
231 return -EINVAL;
232 }
233
234 dbs_data->up_threshold = input;
235 return count;
236}
237
238static ssize_t store_sampling_down_factor(struct gov_attr_set *attr_set,
239 const char *buf, size_t count)
240{
241 struct dbs_data *dbs_data = to_dbs_data(attr_set);
242 struct policy_dbs_info *policy_dbs;
243 unsigned int input;
244 int ret;
245 ret = sscanf(buf, "%u", &input);
246
247 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
248 return -EINVAL;
249
250 dbs_data->sampling_down_factor = input;
251
252 /* Reset down sampling multiplier in case it was active */
253 list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
254 /*
255 * Doing this without locking might lead to using different
256 * rate_mult values in od_update() and od_dbs_update().
257 */
258 mutex_lock(&policy_dbs->update_mutex);
259 policy_dbs->rate_mult = 1;
260 mutex_unlock(&policy_dbs->update_mutex);
261 }
262
263 return count;
264}
265
266static ssize_t store_ignore_nice_load(struct gov_attr_set *attr_set,
267 const char *buf, size_t count)
268{
269 struct dbs_data *dbs_data = to_dbs_data(attr_set);
270 unsigned int input;
271 int ret;
272
273 ret = sscanf(buf, "%u", &input);
274 if (ret != 1)
275 return -EINVAL;
276
277 if (input > 1)
278 input = 1;
279
280 if (input == dbs_data->ignore_nice_load) { /* nothing to do */
281 return count;
282 }
283 dbs_data->ignore_nice_load = input;
284
285 /* we need to re-evaluate prev_cpu_idle */
286 gov_update_cpu_data(dbs_data);
287
288 return count;
289}
290
291static ssize_t store_powersave_bias(struct gov_attr_set *attr_set,
292 const char *buf, size_t count)
293{
294 struct dbs_data *dbs_data = to_dbs_data(attr_set);
295 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
296 struct policy_dbs_info *policy_dbs;
297 unsigned int input;
298 int ret;
299 ret = sscanf(buf, "%u", &input);
300
301 if (ret != 1)
302 return -EINVAL;
303
304 if (input > 1000)
305 input = 1000;
306
307 od_tuners->powersave_bias = input;
308
309 list_for_each_entry(policy_dbs, &attr_set->policy_list, list)
310 ondemand_powersave_bias_init(policy_dbs->policy);
311
312 return count;
313}
314
315gov_show_one_common(sampling_rate);
316gov_show_one_common(up_threshold);
317gov_show_one_common(sampling_down_factor);
318gov_show_one_common(ignore_nice_load);
319gov_show_one_common(io_is_busy);
320gov_show_one(od, powersave_bias);
321
322gov_attr_rw(sampling_rate);
323gov_attr_rw(io_is_busy);
324gov_attr_rw(up_threshold);
325gov_attr_rw(sampling_down_factor);
326gov_attr_rw(ignore_nice_load);
327gov_attr_rw(powersave_bias);
328
329static struct attribute *od_attributes[] = {
330 &sampling_rate.attr,
331 &up_threshold.attr,
332 &sampling_down_factor.attr,
333 &ignore_nice_load.attr,
334 &powersave_bias.attr,
335 &io_is_busy.attr,
336 NULL
337};
338
339/************************** sysfs end ************************/
340
341static struct policy_dbs_info *od_alloc(void)
342{
343 struct od_policy_dbs_info *dbs_info;
344
345 dbs_info = kzalloc(sizeof(*dbs_info), GFP_KERNEL);
346 return dbs_info ? &dbs_info->policy_dbs : NULL;
347}
348
349static void od_free(struct policy_dbs_info *policy_dbs)
350{
351 kfree(to_dbs_info(policy_dbs));
352}
353
354static int od_init(struct dbs_data *dbs_data)
355{
356 struct od_dbs_tuners *tuners;
357 u64 idle_time;
358 int cpu;
359
360 tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
361 if (!tuners)
362 return -ENOMEM;
363
364 cpu = get_cpu();
365 idle_time = get_cpu_idle_time_us(cpu, NULL);
366 put_cpu();
367 if (idle_time != -1ULL) {
368 /* Idle micro accounting is supported. Use finer thresholds */
369 dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
370 } else {
371 dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
372 }
373
374 dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
375 dbs_data->ignore_nice_load = 0;
376 tuners->powersave_bias = default_powersave_bias;
377 dbs_data->io_is_busy = should_io_be_busy();
378
379 dbs_data->tuners = tuners;
380 return 0;
381}
382
383static void od_exit(struct dbs_data *dbs_data)
384{
385 kfree(dbs_data->tuners);
386}
387
388static void od_start(struct cpufreq_policy *policy)
389{
390 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
391
392 dbs_info->sample_type = OD_NORMAL_SAMPLE;
393 ondemand_powersave_bias_init(policy);
394}
395
396static struct od_ops od_ops = {
397 .powersave_bias_target = generic_powersave_bias_target,
398};
399
400static struct dbs_governor od_dbs_gov = {
401 .gov = CPUFREQ_DBS_GOVERNOR_INITIALIZER("ondemand"),
402 .kobj_type = { .default_attrs = od_attributes },
403 .gov_dbs_update = od_dbs_update,
404 .alloc = od_alloc,
405 .free = od_free,
406 .init = od_init,
407 .exit = od_exit,
408 .start = od_start,
409};
410
411#define CPU_FREQ_GOV_ONDEMAND (od_dbs_gov.gov)
412
413static void od_set_powersave_bias(unsigned int powersave_bias)
414{
415 unsigned int cpu;
416 cpumask_t done;
417
418 default_powersave_bias = powersave_bias;
419 cpumask_clear(&done);
420
421 get_online_cpus();
422 for_each_online_cpu(cpu) {
423 struct cpufreq_policy *policy;
424 struct policy_dbs_info *policy_dbs;
425 struct dbs_data *dbs_data;
426 struct od_dbs_tuners *od_tuners;
427
428 if (cpumask_test_cpu(cpu, &done))
429 continue;
430
431 policy = cpufreq_cpu_get_raw(cpu);
432 if (!policy || policy->governor != &CPU_FREQ_GOV_ONDEMAND)
433 continue;
434
435 policy_dbs = policy->governor_data;
436 if (!policy_dbs)
437 continue;
438
439 cpumask_or(&done, &done, policy->cpus);
440
441 dbs_data = policy_dbs->dbs_data;
442 od_tuners = dbs_data->tuners;
443 od_tuners->powersave_bias = default_powersave_bias;
444 }
445 put_online_cpus();
446}
447
448void od_register_powersave_bias_handler(unsigned int (*f)
449 (struct cpufreq_policy *, unsigned int, unsigned int),
450 unsigned int powersave_bias)
451{
452 od_ops.powersave_bias_target = f;
453 od_set_powersave_bias(powersave_bias);
454}
455EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
456
457void od_unregister_powersave_bias_handler(void)
458{
459 od_ops.powersave_bias_target = generic_powersave_bias_target;
460 od_set_powersave_bias(0);
461}
462EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
463
464MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
465MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
466MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
467 "Low Latency Frequency Transition capable processors");
468MODULE_LICENSE("GPL");
469
470#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
471struct cpufreq_governor *cpufreq_default_governor(void)
472{
473 return &CPU_FREQ_GOV_ONDEMAND;
474}
475#endif
476
477cpufreq_governor_init(CPU_FREQ_GOV_ONDEMAND);
478cpufreq_governor_exit(CPU_FREQ_GOV_ONDEMAND);