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
);
173 /* Calculate the next frequency proportional to load */
174 unsigned int freq_next
;
175 freq_next
= load
* policy
->cpuinfo
.max_freq
/ 100;
177 /* No longer fully busy, reset rate_mult */
178 dbs_info
->rate_mult
= 1;
180 if (freq_next
< policy
->min
)
181 freq_next
= policy
->min
;
183 if (!od_tuners
->powersave_bias
) {
184 __cpufreq_driver_target(policy
, freq_next
,
189 freq_next
= od_ops
.powersave_bias_target(policy
, freq_next
,
191 __cpufreq_driver_target(policy
, freq_next
, CPUFREQ_RELATION_L
);
195 static void od_dbs_timer(struct work_struct
*work
)
197 struct od_cpu_dbs_info_s
*dbs_info
=
198 container_of(work
, struct od_cpu_dbs_info_s
, cdbs
.work
.work
);
199 unsigned int cpu
= dbs_info
->cdbs
.cur_policy
->cpu
;
200 struct od_cpu_dbs_info_s
*core_dbs_info
= &per_cpu(od_cpu_dbs_info
,
202 struct dbs_data
*dbs_data
= dbs_info
->cdbs
.cur_policy
->governor_data
;
203 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
204 int delay
= 0, sample_type
= core_dbs_info
->sample_type
;
205 bool modify_all
= true;
207 mutex_lock(&core_dbs_info
->cdbs
.timer_mutex
);
208 if (!need_load_eval(&core_dbs_info
->cdbs
, od_tuners
->sampling_rate
)) {
213 /* Common NORMAL_SAMPLE setup */
214 core_dbs_info
->sample_type
= OD_NORMAL_SAMPLE
;
215 if (sample_type
== OD_SUB_SAMPLE
) {
216 delay
= core_dbs_info
->freq_lo_jiffies
;
217 __cpufreq_driver_target(core_dbs_info
->cdbs
.cur_policy
,
218 core_dbs_info
->freq_lo
, CPUFREQ_RELATION_H
);
220 dbs_check_cpu(dbs_data
, cpu
);
221 if (core_dbs_info
->freq_lo
) {
222 /* Setup timer for SUB_SAMPLE */
223 core_dbs_info
->sample_type
= OD_SUB_SAMPLE
;
224 delay
= core_dbs_info
->freq_hi_jiffies
;
230 delay
= delay_for_sampling_rate(od_tuners
->sampling_rate
231 * core_dbs_info
->rate_mult
);
233 gov_queue_work(dbs_data
, dbs_info
->cdbs
.cur_policy
, delay
, modify_all
);
234 mutex_unlock(&core_dbs_info
->cdbs
.timer_mutex
);
237 /************************** sysfs interface ************************/
238 static struct common_dbs_data od_dbs_cdata
;
241 * update_sampling_rate - update sampling rate effective immediately if needed.
242 * @new_rate: new sampling rate
244 * If new rate is smaller than the old, simply updating
245 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
246 * original sampling_rate was 1 second and the requested new sampling rate is 10
247 * ms because the user needs immediate reaction from ondemand governor, but not
248 * sure if higher frequency will be required or not, then, the governor may
249 * change the sampling rate too late; up to 1 second later. Thus, if we are
250 * reducing the sampling rate, we need to make the new value effective
253 static void update_sampling_rate(struct dbs_data
*dbs_data
,
254 unsigned int new_rate
)
256 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
259 od_tuners
->sampling_rate
= new_rate
= max(new_rate
,
260 dbs_data
->min_sampling_rate
);
262 for_each_online_cpu(cpu
) {
263 struct cpufreq_policy
*policy
;
264 struct od_cpu_dbs_info_s
*dbs_info
;
265 unsigned long next_sampling
, appointed_at
;
267 policy
= cpufreq_cpu_get(cpu
);
270 if (policy
->governor
!= &cpufreq_gov_ondemand
) {
271 cpufreq_cpu_put(policy
);
274 dbs_info
= &per_cpu(od_cpu_dbs_info
, cpu
);
275 cpufreq_cpu_put(policy
);
277 mutex_lock(&dbs_info
->cdbs
.timer_mutex
);
279 if (!delayed_work_pending(&dbs_info
->cdbs
.work
)) {
280 mutex_unlock(&dbs_info
->cdbs
.timer_mutex
);
284 next_sampling
= jiffies
+ usecs_to_jiffies(new_rate
);
285 appointed_at
= dbs_info
->cdbs
.work
.timer
.expires
;
287 if (time_before(next_sampling
, appointed_at
)) {
289 mutex_unlock(&dbs_info
->cdbs
.timer_mutex
);
290 cancel_delayed_work_sync(&dbs_info
->cdbs
.work
);
291 mutex_lock(&dbs_info
->cdbs
.timer_mutex
);
293 gov_queue_work(dbs_data
, dbs_info
->cdbs
.cur_policy
,
294 usecs_to_jiffies(new_rate
), true);
297 mutex_unlock(&dbs_info
->cdbs
.timer_mutex
);
301 static ssize_t
store_sampling_rate(struct dbs_data
*dbs_data
, const char *buf
,
306 ret
= sscanf(buf
, "%u", &input
);
310 update_sampling_rate(dbs_data
, input
);
314 static ssize_t
store_io_is_busy(struct dbs_data
*dbs_data
, const char *buf
,
317 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
322 ret
= sscanf(buf
, "%u", &input
);
325 od_tuners
->io_is_busy
= !!input
;
327 /* we need to re-evaluate prev_cpu_idle */
328 for_each_online_cpu(j
) {
329 struct od_cpu_dbs_info_s
*dbs_info
= &per_cpu(od_cpu_dbs_info
,
331 dbs_info
->cdbs
.prev_cpu_idle
= get_cpu_idle_time(j
,
332 &dbs_info
->cdbs
.prev_cpu_wall
, od_tuners
->io_is_busy
);
337 static ssize_t
store_up_threshold(struct dbs_data
*dbs_data
, const char *buf
,
340 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
343 ret
= sscanf(buf
, "%u", &input
);
345 if (ret
!= 1 || input
> MAX_FREQUENCY_UP_THRESHOLD
||
346 input
< MIN_FREQUENCY_UP_THRESHOLD
) {
350 od_tuners
->up_threshold
= input
;
354 static ssize_t
store_sampling_down_factor(struct dbs_data
*dbs_data
,
355 const char *buf
, size_t count
)
357 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
358 unsigned int input
, j
;
360 ret
= sscanf(buf
, "%u", &input
);
362 if (ret
!= 1 || input
> MAX_SAMPLING_DOWN_FACTOR
|| input
< 1)
364 od_tuners
->sampling_down_factor
= input
;
366 /* Reset down sampling multiplier in case it was active */
367 for_each_online_cpu(j
) {
368 struct od_cpu_dbs_info_s
*dbs_info
= &per_cpu(od_cpu_dbs_info
,
370 dbs_info
->rate_mult
= 1;
375 static ssize_t
store_ignore_nice(struct dbs_data
*dbs_data
, const char *buf
,
378 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
384 ret
= sscanf(buf
, "%u", &input
);
391 if (input
== od_tuners
->ignore_nice
) { /* nothing to do */
394 od_tuners
->ignore_nice
= input
;
396 /* we need to re-evaluate prev_cpu_idle */
397 for_each_online_cpu(j
) {
398 struct od_cpu_dbs_info_s
*dbs_info
;
399 dbs_info
= &per_cpu(od_cpu_dbs_info
, j
);
400 dbs_info
->cdbs
.prev_cpu_idle
= get_cpu_idle_time(j
,
401 &dbs_info
->cdbs
.prev_cpu_wall
, od_tuners
->io_is_busy
);
402 if (od_tuners
->ignore_nice
)
403 dbs_info
->cdbs
.prev_cpu_nice
=
404 kcpustat_cpu(j
).cpustat
[CPUTIME_NICE
];
410 static ssize_t
store_powersave_bias(struct dbs_data
*dbs_data
, const char *buf
,
413 struct od_dbs_tuners
*od_tuners
= dbs_data
->tuners
;
416 ret
= sscanf(buf
, "%u", &input
);
424 od_tuners
->powersave_bias
= input
;
425 ondemand_powersave_bias_init();
429 show_store_one(od
, sampling_rate
);
430 show_store_one(od
, io_is_busy
);
431 show_store_one(od
, up_threshold
);
432 show_store_one(od
, sampling_down_factor
);
433 show_store_one(od
, ignore_nice
);
434 show_store_one(od
, powersave_bias
);
435 declare_show_sampling_rate_min(od
);
437 gov_sys_pol_attr_rw(sampling_rate
);
438 gov_sys_pol_attr_rw(io_is_busy
);
439 gov_sys_pol_attr_rw(up_threshold
);
440 gov_sys_pol_attr_rw(sampling_down_factor
);
441 gov_sys_pol_attr_rw(ignore_nice
);
442 gov_sys_pol_attr_rw(powersave_bias
);
443 gov_sys_pol_attr_ro(sampling_rate_min
);
445 static struct attribute
*dbs_attributes_gov_sys
[] = {
446 &sampling_rate_min_gov_sys
.attr
,
447 &sampling_rate_gov_sys
.attr
,
448 &up_threshold_gov_sys
.attr
,
449 &sampling_down_factor_gov_sys
.attr
,
450 &ignore_nice_gov_sys
.attr
,
451 &powersave_bias_gov_sys
.attr
,
452 &io_is_busy_gov_sys
.attr
,
456 static struct attribute_group od_attr_group_gov_sys
= {
457 .attrs
= dbs_attributes_gov_sys
,
461 static struct attribute
*dbs_attributes_gov_pol
[] = {
462 &sampling_rate_min_gov_pol
.attr
,
463 &sampling_rate_gov_pol
.attr
,
464 &up_threshold_gov_pol
.attr
,
465 &sampling_down_factor_gov_pol
.attr
,
466 &ignore_nice_gov_pol
.attr
,
467 &powersave_bias_gov_pol
.attr
,
468 &io_is_busy_gov_pol
.attr
,
472 static struct attribute_group od_attr_group_gov_pol
= {
473 .attrs
= dbs_attributes_gov_pol
,
477 /************************** sysfs end ************************/
479 static int od_init(struct dbs_data
*dbs_data
)
481 struct od_dbs_tuners
*tuners
;
485 tuners
= kzalloc(sizeof(*tuners
), GFP_KERNEL
);
487 pr_err("%s: kzalloc failed\n", __func__
);
492 idle_time
= get_cpu_idle_time_us(cpu
, NULL
);
494 if (idle_time
!= -1ULL) {
495 /* Idle micro accounting is supported. Use finer thresholds */
496 tuners
->up_threshold
= MICRO_FREQUENCY_UP_THRESHOLD
;
498 * In nohz/micro accounting case we set the minimum frequency
499 * not depending on HZ, but fixed (very low). The deferred
500 * timer might skip some samples if idle/sleeping as needed.
502 dbs_data
->min_sampling_rate
= MICRO_FREQUENCY_MIN_SAMPLE_RATE
;
504 tuners
->up_threshold
= DEF_FREQUENCY_UP_THRESHOLD
;
506 /* For correct statistics, we need 10 ticks for each measure */
507 dbs_data
->min_sampling_rate
= MIN_SAMPLING_RATE_RATIO
*
508 jiffies_to_usecs(10);
511 tuners
->sampling_down_factor
= DEF_SAMPLING_DOWN_FACTOR
;
512 tuners
->ignore_nice
= 0;
513 tuners
->powersave_bias
= default_powersave_bias
;
514 tuners
->io_is_busy
= should_io_be_busy();
516 dbs_data
->tuners
= tuners
;
517 mutex_init(&dbs_data
->mutex
);
521 static void od_exit(struct dbs_data
*dbs_data
)
523 kfree(dbs_data
->tuners
);
526 define_get_cpu_dbs_routines(od_cpu_dbs_info
);
528 static struct od_ops od_ops
= {
529 .powersave_bias_init_cpu
= ondemand_powersave_bias_init_cpu
,
530 .powersave_bias_target
= generic_powersave_bias_target
,
531 .freq_increase
= dbs_freq_increase
,
534 static struct common_dbs_data od_dbs_cdata
= {
535 .governor
= GOV_ONDEMAND
,
536 .attr_group_gov_sys
= &od_attr_group_gov_sys
,
537 .attr_group_gov_pol
= &od_attr_group_gov_pol
,
538 .get_cpu_cdbs
= get_cpu_cdbs
,
539 .get_cpu_dbs_info_s
= get_cpu_dbs_info_s
,
540 .gov_dbs_timer
= od_dbs_timer
,
541 .gov_check_cpu
= od_check_cpu
,
547 static void od_set_powersave_bias(unsigned int powersave_bias
)
549 struct cpufreq_policy
*policy
;
550 struct dbs_data
*dbs_data
;
551 struct od_dbs_tuners
*od_tuners
;
555 default_powersave_bias
= powersave_bias
;
556 cpumask_clear(&done
);
559 for_each_online_cpu(cpu
) {
560 if (cpumask_test_cpu(cpu
, &done
))
563 policy
= per_cpu(od_cpu_dbs_info
, cpu
).cdbs
.cur_policy
;
567 cpumask_or(&done
, &done
, policy
->cpus
);
569 if (policy
->governor
!= &cpufreq_gov_ondemand
)
572 dbs_data
= policy
->governor_data
;
573 od_tuners
= dbs_data
->tuners
;
574 od_tuners
->powersave_bias
= default_powersave_bias
;
579 void od_register_powersave_bias_handler(unsigned int (*f
)
580 (struct cpufreq_policy
*, unsigned int, unsigned int),
581 unsigned int powersave_bias
)
583 od_ops
.powersave_bias_target
= f
;
584 od_set_powersave_bias(powersave_bias
);
586 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler
);
588 void od_unregister_powersave_bias_handler(void)
590 od_ops
.powersave_bias_target
= generic_powersave_bias_target
;
591 od_set_powersave_bias(0);
593 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler
);
595 static int od_cpufreq_governor_dbs(struct cpufreq_policy
*policy
,
598 return cpufreq_governor_dbs(policy
, &od_dbs_cdata
, event
);
601 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
604 struct cpufreq_governor cpufreq_gov_ondemand
= {
606 .governor
= od_cpufreq_governor_dbs
,
607 .max_transition_latency
= TRANSITION_LATENCY_LIMIT
,
608 .owner
= THIS_MODULE
,
611 static int __init
cpufreq_gov_dbs_init(void)
613 return cpufreq_register_governor(&cpufreq_gov_ondemand
);
616 static void __exit
cpufreq_gov_dbs_exit(void)
618 cpufreq_unregister_governor(&cpufreq_gov_ondemand
);
621 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
622 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
623 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
624 "Low Latency Frequency Transition capable processors");
625 MODULE_LICENSE("GPL");
627 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
628 fs_initcall(cpufreq_gov_dbs_init
);
630 module_init(cpufreq_gov_dbs_init
);
632 module_exit(cpufreq_gov_dbs_exit
);