2 * drivers/cpufreq/cpufreq_ondemand.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
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.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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"
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)
30 static DEFINE_PER_CPU(struct od_cpu_dbs_info_s
, od_cpu_dbs_info
);
32 static struct od_ops od_ops
;
34 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
35 static struct cpufreq_governor cpufreq_gov_ondemand
;
38 static unsigned int default_powersave_bias
;
40 static void ondemand_powersave_bias_init_cpu(int cpu
)
42 struct od_cpu_dbs_info_s
*dbs_info
= &per_cpu(od_cpu_dbs_info
, cpu
);
44 dbs_info
->freq_table
= cpufreq_frequency_get_table(cpu
);
45 dbs_info
->freq_lo
= 0;
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
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.
57 static int should_io_be_busy(void)
59 #if defined(CONFIG_X86)
61 * For Intel, Core 2 (model 15) and later have an efficient idle.
63 if (boot_cpu_data
.x86_vendor
== X86_VENDOR_INTEL
&&
64 boot_cpu_data
.x86
== 6 &&
65 boot_cpu_data
.x86_model
>= 15)
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.
76 static unsigned int generic_powersave_bias_target(struct cpufreq_policy
*policy
,
77 unsigned int freq_next
, unsigned int relation
)
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
,
85 struct dbs_data
*dbs_data
= policy
->governor_data
;
86 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
88 if (!dbs_info
->freq_table
) {
89 dbs_info
->freq_lo
= 0;
90 dbs_info
->freq_lo_jiffies
= 0;
94 cpufreq_frequency_table_target(policy
, dbs_info
->freq_table
, freq_next
,
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
;
100 /* Find freq bounds for freq_avg in freq_table */
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
;
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
;
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;
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
;
127 static void ondemand_powersave_bias_init(void)
130 for_each_online_cpu(i
) {
131 ondemand_powersave_bias_init_cpu(i
);
135 static void dbs_freq_increase(struct cpufreq_policy
*policy
, unsigned int freq
)
137 struct dbs_data
*dbs_data
= policy
->governor_data
;
138 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
140 if (od_tuners
->powersave_bias
)
141 freq
= od_ops
.powersave_bias_target(policy
, freq
,
143 else if (policy
->cur
== policy
->max
)
146 __cpufreq_driver_target(policy
, freq
, od_tuners
->powersave_bias
?
147 CPUFREQ_RELATION_L
: CPUFREQ_RELATION_H
);
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.
155 static void od_check_cpu(int cpu
, unsigned int load
)
157 struct od_cpu_dbs_info_s
*dbs_info
= &per_cpu(od_cpu_dbs_info
, cpu
);
158 struct cpufreq_policy
*policy
= dbs_info
->cdbs
.cur_policy
;
159 struct dbs_data
*dbs_data
= policy
->governor_data
;
160 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
162 dbs_info
->freq_lo
= 0;
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
);
172 /* Calculate the next frequency proportional to load */
173 unsigned int freq_next
, min_f
, max_f
;
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;
179 /* No longer fully busy, reset rate_mult */
180 dbs_info
->rate_mult
= 1;
182 if (!od_tuners
->powersave_bias
) {
183 __cpufreq_driver_target(policy
, freq_next
,
188 freq_next
= od_ops
.powersave_bias_target(policy
, freq_next
,
190 __cpufreq_driver_target(policy
, freq_next
, CPUFREQ_RELATION_C
);
194 static void od_dbs_timer(struct work_struct
*work
)
196 struct od_cpu_dbs_info_s
*dbs_info
=
197 container_of(work
, struct od_cpu_dbs_info_s
, cdbs
.work
.work
);
198 unsigned int cpu
= dbs_info
->cdbs
.cur_policy
->cpu
;
199 struct od_cpu_dbs_info_s
*core_dbs_info
= &per_cpu(od_cpu_dbs_info
,
201 struct dbs_data
*dbs_data
= dbs_info
->cdbs
.cur_policy
->governor_data
;
202 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
203 int delay
= 0, sample_type
= core_dbs_info
->sample_type
;
204 bool modify_all
= true;
206 mutex_lock(&core_dbs_info
->cdbs
.timer_mutex
);
207 if (!need_load_eval(&core_dbs_info
->cdbs
, od_tuners
->sampling_rate
)) {
212 /* Common NORMAL_SAMPLE setup */
213 core_dbs_info
->sample_type
= OD_NORMAL_SAMPLE
;
214 if (sample_type
== OD_SUB_SAMPLE
) {
215 delay
= core_dbs_info
->freq_lo_jiffies
;
216 __cpufreq_driver_target(core_dbs_info
->cdbs
.cur_policy
,
217 core_dbs_info
->freq_lo
, CPUFREQ_RELATION_H
);
219 dbs_check_cpu(dbs_data
, cpu
);
220 if (core_dbs_info
->freq_lo
) {
221 /* Setup timer for SUB_SAMPLE */
222 core_dbs_info
->sample_type
= OD_SUB_SAMPLE
;
223 delay
= core_dbs_info
->freq_hi_jiffies
;
229 delay
= delay_for_sampling_rate(od_tuners
->sampling_rate
230 * core_dbs_info
->rate_mult
);
232 gov_queue_work(dbs_data
, dbs_info
->cdbs
.cur_policy
, delay
, modify_all
);
233 mutex_unlock(&core_dbs_info
->cdbs
.timer_mutex
);
236 /************************** sysfs interface ************************/
237 static struct common_dbs_data od_dbs_cdata
;
240 * update_sampling_rate - update sampling rate effective immediately if needed.
241 * @new_rate: new sampling rate
243 * If new rate is smaller than the old, simply updating
244 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
245 * original sampling_rate was 1 second and the requested new sampling rate is 10
246 * ms because the user needs immediate reaction from ondemand governor, but not
247 * sure if higher frequency will be required or not, then, the governor may
248 * change the sampling rate too late; up to 1 second later. Thus, if we are
249 * reducing the sampling rate, we need to make the new value effective
252 static void update_sampling_rate(struct dbs_data
*dbs_data
,
253 unsigned int new_rate
)
255 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
258 od_tuners
->sampling_rate
= new_rate
= max(new_rate
,
259 dbs_data
->min_sampling_rate
);
261 for_each_online_cpu(cpu
) {
262 struct cpufreq_policy
*policy
;
263 struct od_cpu_dbs_info_s
*dbs_info
;
264 unsigned long next_sampling
, appointed_at
;
266 policy
= cpufreq_cpu_get(cpu
);
269 if (policy
->governor
!= &cpufreq_gov_ondemand
) {
270 cpufreq_cpu_put(policy
);
273 dbs_info
= &per_cpu(od_cpu_dbs_info
, cpu
);
274 cpufreq_cpu_put(policy
);
276 mutex_lock(&dbs_info
->cdbs
.timer_mutex
);
278 if (!delayed_work_pending(&dbs_info
->cdbs
.work
)) {
279 mutex_unlock(&dbs_info
->cdbs
.timer_mutex
);
283 next_sampling
= jiffies
+ usecs_to_jiffies(new_rate
);
284 appointed_at
= dbs_info
->cdbs
.work
.timer
.expires
;
286 if (time_before(next_sampling
, appointed_at
)) {
288 mutex_unlock(&dbs_info
->cdbs
.timer_mutex
);
289 cancel_delayed_work_sync(&dbs_info
->cdbs
.work
);
290 mutex_lock(&dbs_info
->cdbs
.timer_mutex
);
292 gov_queue_work(dbs_data
, dbs_info
->cdbs
.cur_policy
,
293 usecs_to_jiffies(new_rate
), true);
296 mutex_unlock(&dbs_info
->cdbs
.timer_mutex
);
300 static ssize_t
store_sampling_rate(struct dbs_data
*dbs_data
, const char *buf
,
305 ret
= sscanf(buf
, "%u", &input
);
309 update_sampling_rate(dbs_data
, input
);
313 static ssize_t
store_io_is_busy(struct dbs_data
*dbs_data
, const char *buf
,
316 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
321 ret
= sscanf(buf
, "%u", &input
);
324 od_tuners
->io_is_busy
= !!input
;
326 /* we need to re-evaluate prev_cpu_idle */
327 for_each_online_cpu(j
) {
328 struct od_cpu_dbs_info_s
*dbs_info
= &per_cpu(od_cpu_dbs_info
,
330 dbs_info
->cdbs
.prev_cpu_idle
= get_cpu_idle_time(j
,
331 &dbs_info
->cdbs
.prev_cpu_wall
, od_tuners
->io_is_busy
);
336 static ssize_t
store_up_threshold(struct dbs_data
*dbs_data
, const char *buf
,
339 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
342 ret
= sscanf(buf
, "%u", &input
);
344 if (ret
!= 1 || input
> MAX_FREQUENCY_UP_THRESHOLD
||
345 input
< MIN_FREQUENCY_UP_THRESHOLD
) {
349 od_tuners
->up_threshold
= input
;
353 static ssize_t
store_sampling_down_factor(struct dbs_data
*dbs_data
,
354 const char *buf
, size_t count
)
356 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
357 unsigned int input
, j
;
359 ret
= sscanf(buf
, "%u", &input
);
361 if (ret
!= 1 || input
> MAX_SAMPLING_DOWN_FACTOR
|| input
< 1)
363 od_tuners
->sampling_down_factor
= input
;
365 /* Reset down sampling multiplier in case it was active */
366 for_each_online_cpu(j
) {
367 struct od_cpu_dbs_info_s
*dbs_info
= &per_cpu(od_cpu_dbs_info
,
369 dbs_info
->rate_mult
= 1;
374 static ssize_t
store_ignore_nice_load(struct dbs_data
*dbs_data
,
375 const char *buf
, size_t count
)
377 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
383 ret
= sscanf(buf
, "%u", &input
);
390 if (input
== od_tuners
->ignore_nice_load
) { /* nothing to do */
393 od_tuners
->ignore_nice_load
= input
;
395 /* we need to re-evaluate prev_cpu_idle */
396 for_each_online_cpu(j
) {
397 struct od_cpu_dbs_info_s
*dbs_info
;
398 dbs_info
= &per_cpu(od_cpu_dbs_info
, j
);
399 dbs_info
->cdbs
.prev_cpu_idle
= get_cpu_idle_time(j
,
400 &dbs_info
->cdbs
.prev_cpu_wall
, od_tuners
->io_is_busy
);
401 if (od_tuners
->ignore_nice_load
)
402 dbs_info
->cdbs
.prev_cpu_nice
=
403 kcpustat_cpu(j
).cpustat
[CPUTIME_NICE
];
409 static ssize_t
store_powersave_bias(struct dbs_data
*dbs_data
, const char *buf
,
412 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
415 ret
= sscanf(buf
, "%u", &input
);
423 od_tuners
->powersave_bias
= input
;
424 ondemand_powersave_bias_init();
428 show_store_one(od
, sampling_rate
);
429 show_store_one(od
, io_is_busy
);
430 show_store_one(od
, up_threshold
);
431 show_store_one(od
, sampling_down_factor
);
432 show_store_one(od
, ignore_nice_load
);
433 show_store_one(od
, powersave_bias
);
434 declare_show_sampling_rate_min(od
);
436 gov_sys_pol_attr_rw(sampling_rate
);
437 gov_sys_pol_attr_rw(io_is_busy
);
438 gov_sys_pol_attr_rw(up_threshold
);
439 gov_sys_pol_attr_rw(sampling_down_factor
);
440 gov_sys_pol_attr_rw(ignore_nice_load
);
441 gov_sys_pol_attr_rw(powersave_bias
);
442 gov_sys_pol_attr_ro(sampling_rate_min
);
444 static struct attribute
*dbs_attributes_gov_sys
[] = {
445 &sampling_rate_min_gov_sys
.attr
,
446 &sampling_rate_gov_sys
.attr
,
447 &up_threshold_gov_sys
.attr
,
448 &sampling_down_factor_gov_sys
.attr
,
449 &ignore_nice_load_gov_sys
.attr
,
450 &powersave_bias_gov_sys
.attr
,
451 &io_is_busy_gov_sys
.attr
,
455 static struct attribute_group od_attr_group_gov_sys
= {
456 .attrs
= dbs_attributes_gov_sys
,
460 static struct attribute
*dbs_attributes_gov_pol
[] = {
461 &sampling_rate_min_gov_pol
.attr
,
462 &sampling_rate_gov_pol
.attr
,
463 &up_threshold_gov_pol
.attr
,
464 &sampling_down_factor_gov_pol
.attr
,
465 &ignore_nice_load_gov_pol
.attr
,
466 &powersave_bias_gov_pol
.attr
,
467 &io_is_busy_gov_pol
.attr
,
471 static struct attribute_group od_attr_group_gov_pol
= {
472 .attrs
= dbs_attributes_gov_pol
,
476 /************************** sysfs end ************************/
478 static int od_init(struct dbs_data
*dbs_data
)
480 struct od_dbs_tuners
*tuners
;
484 tuners
= kzalloc(sizeof(*tuners
), GFP_KERNEL
);
486 pr_err("%s: kzalloc failed\n", __func__
);
491 idle_time
= get_cpu_idle_time_us(cpu
, NULL
);
493 if (idle_time
!= -1ULL) {
494 /* Idle micro accounting is supported. Use finer thresholds */
495 tuners
->up_threshold
= MICRO_FREQUENCY_UP_THRESHOLD
;
497 * In nohz/micro accounting case we set the minimum frequency
498 * not depending on HZ, but fixed (very low). The deferred
499 * timer might skip some samples if idle/sleeping as needed.
501 dbs_data
->min_sampling_rate
= MICRO_FREQUENCY_MIN_SAMPLE_RATE
;
503 tuners
->up_threshold
= DEF_FREQUENCY_UP_THRESHOLD
;
505 /* For correct statistics, we need 10 ticks for each measure */
506 dbs_data
->min_sampling_rate
= MIN_SAMPLING_RATE_RATIO
*
507 jiffies_to_usecs(10);
510 tuners
->sampling_down_factor
= DEF_SAMPLING_DOWN_FACTOR
;
511 tuners
->ignore_nice_load
= 0;
512 tuners
->powersave_bias
= default_powersave_bias
;
513 tuners
->io_is_busy
= should_io_be_busy();
515 dbs_data
->tuners
= tuners
;
516 mutex_init(&dbs_data
->mutex
);
520 static void od_exit(struct dbs_data
*dbs_data
)
522 kfree(dbs_data
->tuners
);
525 define_get_cpu_dbs_routines(od_cpu_dbs_info
);
527 static struct od_ops od_ops
= {
528 .powersave_bias_init_cpu
= ondemand_powersave_bias_init_cpu
,
529 .powersave_bias_target
= generic_powersave_bias_target
,
530 .freq_increase
= dbs_freq_increase
,
533 static struct common_dbs_data od_dbs_cdata
= {
534 .governor
= GOV_ONDEMAND
,
535 .attr_group_gov_sys
= &od_attr_group_gov_sys
,
536 .attr_group_gov_pol
= &od_attr_group_gov_pol
,
537 .get_cpu_cdbs
= get_cpu_cdbs
,
538 .get_cpu_dbs_info_s
= get_cpu_dbs_info_s
,
539 .gov_dbs_timer
= od_dbs_timer
,
540 .gov_check_cpu
= od_check_cpu
,
546 static void od_set_powersave_bias(unsigned int powersave_bias
)
548 struct cpufreq_policy
*policy
;
549 struct dbs_data
*dbs_data
;
550 struct od_dbs_tuners
*od_tuners
;
554 default_powersave_bias
= powersave_bias
;
555 cpumask_clear(&done
);
558 for_each_online_cpu(cpu
) {
559 if (cpumask_test_cpu(cpu
, &done
))
562 policy
= per_cpu(od_cpu_dbs_info
, cpu
).cdbs
.cur_policy
;
566 cpumask_or(&done
, &done
, policy
->cpus
);
568 if (policy
->governor
!= &cpufreq_gov_ondemand
)
571 dbs_data
= policy
->governor_data
;
572 od_tuners
= dbs_data
->tuners
;
573 od_tuners
->powersave_bias
= default_powersave_bias
;
578 void od_register_powersave_bias_handler(unsigned int (*f
)
579 (struct cpufreq_policy
*, unsigned int, unsigned int),
580 unsigned int powersave_bias
)
582 od_ops
.powersave_bias_target
= f
;
583 od_set_powersave_bias(powersave_bias
);
585 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler
);
587 void od_unregister_powersave_bias_handler(void)
589 od_ops
.powersave_bias_target
= generic_powersave_bias_target
;
590 od_set_powersave_bias(0);
592 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler
);
594 static int od_cpufreq_governor_dbs(struct cpufreq_policy
*policy
,
597 return cpufreq_governor_dbs(policy
, &od_dbs_cdata
, event
);
600 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
603 struct cpufreq_governor cpufreq_gov_ondemand
= {
605 .governor
= od_cpufreq_governor_dbs
,
606 .max_transition_latency
= TRANSITION_LATENCY_LIMIT
,
607 .owner
= THIS_MODULE
,
610 static int __init
cpufreq_gov_dbs_init(void)
612 return cpufreq_register_governor(&cpufreq_gov_ondemand
);
615 static void __exit
cpufreq_gov_dbs_exit(void)
617 cpufreq_unregister_governor(&cpufreq_gov_ondemand
);
620 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
621 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
622 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
623 "Low Latency Frequency Transition capable processors");
624 MODULE_LICENSE("GPL");
626 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
627 fs_initcall(cpufreq_gov_dbs_init
);
629 module_init(cpufreq_gov_dbs_init
);
631 module_exit(cpufreq_gov_dbs_exit
);