2 * drivers/cpufreq/cpufreq_conservative.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
7 * (C) 2009 Alexander Clouter <alex@digriz.org.uk>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 as
11 * published by the Free Software Foundation.
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/cpufreq.h>
18 #include <linux/cpu.h>
19 #include <linux/jiffies.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/mutex.h>
22 #include <linux/hrtimer.h>
23 #include <linux/tick.h>
24 #include <linux/ktime.h>
25 #include <linux/sched.h>
28 * dbs is used in this file as a shortform for demandbased switching
29 * It helps to keep variable names smaller, simpler
32 #define DEF_FREQUENCY_UP_THRESHOLD (80)
33 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
36 * The polling frequency of this governor depends on the capability of
37 * the processor. Default polling frequency is 1000 times the transition
38 * latency of the processor. The governor will work on any processor with
39 * transition latency <= 10mS, using appropriate sampling
41 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
42 * this governor will not work.
43 * All times here are in uS.
45 #define MIN_SAMPLING_RATE_RATIO (2)
47 static unsigned int min_sampling_rate
;
49 #define LATENCY_MULTIPLIER (1000)
50 #define MIN_LATENCY_MULTIPLIER (100)
51 #define DEF_SAMPLING_DOWN_FACTOR (1)
52 #define MAX_SAMPLING_DOWN_FACTOR (10)
53 #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
55 static void do_dbs_timer(struct work_struct
*work
);
57 struct cpu_dbs_info_s
{
58 cputime64_t prev_cpu_idle
;
59 cputime64_t prev_cpu_wall
;
60 cputime64_t prev_cpu_nice
;
61 struct cpufreq_policy
*cur_policy
;
62 struct delayed_work work
;
63 unsigned int down_skip
;
64 unsigned int requested_freq
;
66 unsigned int enable
:1;
68 * percpu mutex that serializes governor limit change with
69 * do_dbs_timer invocation. We do not want do_dbs_timer to run
70 * when user is changing the governor or limits.
72 struct mutex timer_mutex
;
74 static DEFINE_PER_CPU(struct cpu_dbs_info_s
, cs_cpu_dbs_info
);
76 static unsigned int dbs_enable
; /* number of CPUs using this policy */
79 * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
80 * different CPUs. It protects dbs_enable in governor start/stop.
82 static DEFINE_MUTEX(dbs_mutex
);
84 static struct dbs_tuners
{
85 unsigned int sampling_rate
;
86 unsigned int sampling_down_factor
;
87 unsigned int up_threshold
;
88 unsigned int down_threshold
;
89 unsigned int ignore_nice
;
90 unsigned int freq_step
;
92 .up_threshold
= DEF_FREQUENCY_UP_THRESHOLD
,
93 .down_threshold
= DEF_FREQUENCY_DOWN_THRESHOLD
,
94 .sampling_down_factor
= DEF_SAMPLING_DOWN_FACTOR
,
99 static inline cputime64_t
get_cpu_idle_time_jiffy(unsigned int cpu
,
102 cputime64_t idle_time
;
103 cputime64_t cur_wall_time
;
104 cputime64_t busy_time
;
106 cur_wall_time
= jiffies64_to_cputime64(get_jiffies_64());
107 busy_time
= cputime64_add(kstat_cpu(cpu
).cpustat
.user
,
108 kstat_cpu(cpu
).cpustat
.system
);
110 busy_time
= cputime64_add(busy_time
, kstat_cpu(cpu
).cpustat
.irq
);
111 busy_time
= cputime64_add(busy_time
, kstat_cpu(cpu
).cpustat
.softirq
);
112 busy_time
= cputime64_add(busy_time
, kstat_cpu(cpu
).cpustat
.steal
);
113 busy_time
= cputime64_add(busy_time
, kstat_cpu(cpu
).cpustat
.nice
);
115 idle_time
= cputime64_sub(cur_wall_time
, busy_time
);
117 *wall
= (cputime64_t
)jiffies_to_usecs(cur_wall_time
);
119 return (cputime64_t
)jiffies_to_usecs(idle_time
);
122 static inline cputime64_t
get_cpu_idle_time(unsigned int cpu
, cputime64_t
*wall
)
124 u64 idle_time
= get_cpu_idle_time_us(cpu
, wall
);
126 if (idle_time
== -1ULL)
127 return get_cpu_idle_time_jiffy(cpu
, wall
);
132 /* keep track of frequency transitions */
134 dbs_cpufreq_notifier(struct notifier_block
*nb
, unsigned long val
,
137 struct cpufreq_freqs
*freq
= data
;
138 struct cpu_dbs_info_s
*this_dbs_info
= &per_cpu(cs_cpu_dbs_info
,
141 struct cpufreq_policy
*policy
;
143 if (!this_dbs_info
->enable
)
146 policy
= this_dbs_info
->cur_policy
;
149 * we only care if our internally tracked freq moves outside
150 * the 'valid' ranges of freqency available to us otherwise
151 * we do not change it
153 if (this_dbs_info
->requested_freq
> policy
->max
154 || this_dbs_info
->requested_freq
< policy
->min
)
155 this_dbs_info
->requested_freq
= freq
->new;
160 static struct notifier_block dbs_cpufreq_notifier_block
= {
161 .notifier_call
= dbs_cpufreq_notifier
164 /************************** sysfs interface ************************/
165 static ssize_t
show_sampling_rate_min(struct kobject
*kobj
,
166 struct attribute
*attr
, char *buf
)
168 return sprintf(buf
, "%u\n", min_sampling_rate
);
171 define_one_global_ro(sampling_rate_min
);
173 /* cpufreq_conservative Governor Tunables */
174 #define show_one(file_name, object) \
175 static ssize_t show_##file_name \
176 (struct kobject *kobj, struct attribute *attr, char *buf) \
178 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
180 show_one(sampling_rate
, sampling_rate
);
181 show_one(sampling_down_factor
, sampling_down_factor
);
182 show_one(up_threshold
, up_threshold
);
183 show_one(down_threshold
, down_threshold
);
184 show_one(ignore_nice_load
, ignore_nice
);
185 show_one(freq_step
, freq_step
);
187 /*** delete after deprecation time ***/
188 #define DEPRECATION_MSG(file_name) \
189 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
190 "interface is deprecated - " #file_name "\n");
192 #define show_one_old(file_name) \
193 static ssize_t show_##file_name##_old \
194 (struct cpufreq_policy *unused, char *buf) \
196 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
197 "interface is deprecated - " #file_name "\n"); \
198 return show_##file_name(NULL, NULL, buf); \
200 show_one_old(sampling_rate
);
201 show_one_old(sampling_down_factor
);
202 show_one_old(up_threshold
);
203 show_one_old(down_threshold
);
204 show_one_old(ignore_nice_load
);
205 show_one_old(freq_step
);
206 show_one_old(sampling_rate_min
);
208 cpufreq_freq_attr_ro_old(sampling_rate_min
);
210 /*** delete after deprecation time ***/
212 static ssize_t
store_sampling_down_factor(struct kobject
*a
,
214 const char *buf
, size_t count
)
218 ret
= sscanf(buf
, "%u", &input
);
220 if (ret
!= 1 || input
> MAX_SAMPLING_DOWN_FACTOR
|| input
< 1)
223 mutex_lock(&dbs_mutex
);
224 dbs_tuners_ins
.sampling_down_factor
= input
;
225 mutex_unlock(&dbs_mutex
);
230 static ssize_t
store_sampling_rate(struct kobject
*a
, struct attribute
*b
,
231 const char *buf
, size_t count
)
235 ret
= sscanf(buf
, "%u", &input
);
240 mutex_lock(&dbs_mutex
);
241 dbs_tuners_ins
.sampling_rate
= max(input
, min_sampling_rate
);
242 mutex_unlock(&dbs_mutex
);
247 static ssize_t
store_up_threshold(struct kobject
*a
, struct attribute
*b
,
248 const char *buf
, size_t count
)
252 ret
= sscanf(buf
, "%u", &input
);
254 mutex_lock(&dbs_mutex
);
255 if (ret
!= 1 || input
> 100 ||
256 input
<= dbs_tuners_ins
.down_threshold
) {
257 mutex_unlock(&dbs_mutex
);
261 dbs_tuners_ins
.up_threshold
= input
;
262 mutex_unlock(&dbs_mutex
);
267 static ssize_t
store_down_threshold(struct kobject
*a
, struct attribute
*b
,
268 const char *buf
, size_t count
)
272 ret
= sscanf(buf
, "%u", &input
);
274 mutex_lock(&dbs_mutex
);
275 /* cannot be lower than 11 otherwise freq will not fall */
276 if (ret
!= 1 || input
< 11 || input
> 100 ||
277 input
>= dbs_tuners_ins
.up_threshold
) {
278 mutex_unlock(&dbs_mutex
);
282 dbs_tuners_ins
.down_threshold
= input
;
283 mutex_unlock(&dbs_mutex
);
288 static ssize_t
store_ignore_nice_load(struct kobject
*a
, struct attribute
*b
,
289 const char *buf
, size_t count
)
296 ret
= sscanf(buf
, "%u", &input
);
303 mutex_lock(&dbs_mutex
);
304 if (input
== dbs_tuners_ins
.ignore_nice
) { /* nothing to do */
305 mutex_unlock(&dbs_mutex
);
308 dbs_tuners_ins
.ignore_nice
= input
;
310 /* we need to re-evaluate prev_cpu_idle */
311 for_each_online_cpu(j
) {
312 struct cpu_dbs_info_s
*dbs_info
;
313 dbs_info
= &per_cpu(cs_cpu_dbs_info
, j
);
314 dbs_info
->prev_cpu_idle
= get_cpu_idle_time(j
,
315 &dbs_info
->prev_cpu_wall
);
316 if (dbs_tuners_ins
.ignore_nice
)
317 dbs_info
->prev_cpu_nice
= kstat_cpu(j
).cpustat
.nice
;
319 mutex_unlock(&dbs_mutex
);
324 static ssize_t
store_freq_step(struct kobject
*a
, struct attribute
*b
,
325 const char *buf
, size_t count
)
329 ret
= sscanf(buf
, "%u", &input
);
337 /* no need to test here if freq_step is zero as the user might actually
338 * want this, they would be crazy though :) */
339 mutex_lock(&dbs_mutex
);
340 dbs_tuners_ins
.freq_step
= input
;
341 mutex_unlock(&dbs_mutex
);
346 define_one_global_rw(sampling_rate
);
347 define_one_global_rw(sampling_down_factor
);
348 define_one_global_rw(up_threshold
);
349 define_one_global_rw(down_threshold
);
350 define_one_global_rw(ignore_nice_load
);
351 define_one_global_rw(freq_step
);
353 static struct attribute
*dbs_attributes
[] = {
354 &sampling_rate_min
.attr
,
356 &sampling_down_factor
.attr
,
358 &down_threshold
.attr
,
359 &ignore_nice_load
.attr
,
364 static struct attribute_group dbs_attr_group
= {
365 .attrs
= dbs_attributes
,
366 .name
= "conservative",
369 /*** delete after deprecation time ***/
371 #define write_one_old(file_name) \
372 static ssize_t store_##file_name##_old \
373 (struct cpufreq_policy *unused, const char *buf, size_t count) \
375 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
376 "interface is deprecated - " #file_name "\n"); \
377 return store_##file_name(NULL, NULL, buf, count); \
379 write_one_old(sampling_rate
);
380 write_one_old(sampling_down_factor
);
381 write_one_old(up_threshold
);
382 write_one_old(down_threshold
);
383 write_one_old(ignore_nice_load
);
384 write_one_old(freq_step
);
386 cpufreq_freq_attr_rw_old(sampling_rate
);
387 cpufreq_freq_attr_rw_old(sampling_down_factor
);
388 cpufreq_freq_attr_rw_old(up_threshold
);
389 cpufreq_freq_attr_rw_old(down_threshold
);
390 cpufreq_freq_attr_rw_old(ignore_nice_load
);
391 cpufreq_freq_attr_rw_old(freq_step
);
393 static struct attribute
*dbs_attributes_old
[] = {
394 &sampling_rate_min_old
.attr
,
395 &sampling_rate_old
.attr
,
396 &sampling_down_factor_old
.attr
,
397 &up_threshold_old
.attr
,
398 &down_threshold_old
.attr
,
399 &ignore_nice_load_old
.attr
,
404 static struct attribute_group dbs_attr_group_old
= {
405 .attrs
= dbs_attributes_old
,
406 .name
= "conservative",
409 /*** delete after deprecation time ***/
411 /************************** sysfs end ************************/
413 static void dbs_check_cpu(struct cpu_dbs_info_s
*this_dbs_info
)
415 unsigned int load
= 0;
416 unsigned int max_load
= 0;
417 unsigned int freq_target
;
419 struct cpufreq_policy
*policy
;
422 policy
= this_dbs_info
->cur_policy
;
425 * Every sampling_rate, we check, if current idle time is less
426 * than 20% (default), then we try to increase frequency
427 * Every sampling_rate*sampling_down_factor, we check, if current
428 * idle time is more than 80%, then we try to decrease frequency
430 * Any frequency increase takes it to the maximum frequency.
431 * Frequency reduction happens at minimum steps of
432 * 5% (default) of maximum frequency
435 /* Get Absolute Load */
436 for_each_cpu(j
, policy
->cpus
) {
437 struct cpu_dbs_info_s
*j_dbs_info
;
438 cputime64_t cur_wall_time
, cur_idle_time
;
439 unsigned int idle_time
, wall_time
;
441 j_dbs_info
= &per_cpu(cs_cpu_dbs_info
, j
);
443 cur_idle_time
= get_cpu_idle_time(j
, &cur_wall_time
);
445 wall_time
= (unsigned int) cputime64_sub(cur_wall_time
,
446 j_dbs_info
->prev_cpu_wall
);
447 j_dbs_info
->prev_cpu_wall
= cur_wall_time
;
449 idle_time
= (unsigned int) cputime64_sub(cur_idle_time
,
450 j_dbs_info
->prev_cpu_idle
);
451 j_dbs_info
->prev_cpu_idle
= cur_idle_time
;
453 if (dbs_tuners_ins
.ignore_nice
) {
454 cputime64_t cur_nice
;
455 unsigned long cur_nice_jiffies
;
457 cur_nice
= cputime64_sub(kstat_cpu(j
).cpustat
.nice
,
458 j_dbs_info
->prev_cpu_nice
);
460 * Assumption: nice time between sampling periods will
461 * be less than 2^32 jiffies for 32 bit sys
463 cur_nice_jiffies
= (unsigned long)
464 cputime64_to_jiffies64(cur_nice
);
466 j_dbs_info
->prev_cpu_nice
= kstat_cpu(j
).cpustat
.nice
;
467 idle_time
+= jiffies_to_usecs(cur_nice_jiffies
);
470 if (unlikely(!wall_time
|| wall_time
< idle_time
))
473 load
= 100 * (wall_time
- idle_time
) / wall_time
;
480 * break out if we 'cannot' reduce the speed as the user might
481 * want freq_step to be zero
483 if (dbs_tuners_ins
.freq_step
== 0)
486 /* Check for frequency increase */
487 if (max_load
> dbs_tuners_ins
.up_threshold
) {
488 this_dbs_info
->down_skip
= 0;
490 /* if we are already at full speed then break out early */
491 if (this_dbs_info
->requested_freq
== policy
->max
)
494 freq_target
= (dbs_tuners_ins
.freq_step
* policy
->max
) / 100;
496 /* max freq cannot be less than 100. But who knows.... */
497 if (unlikely(freq_target
== 0))
500 this_dbs_info
->requested_freq
+= freq_target
;
501 if (this_dbs_info
->requested_freq
> policy
->max
)
502 this_dbs_info
->requested_freq
= policy
->max
;
504 __cpufreq_driver_target(policy
, this_dbs_info
->requested_freq
,
510 * The optimal frequency is the frequency that is the lowest that
511 * can support the current CPU usage without triggering the up
512 * policy. To be safe, we focus 10 points under the threshold.
514 if (max_load
< (dbs_tuners_ins
.down_threshold
- 10)) {
515 freq_target
= (dbs_tuners_ins
.freq_step
* policy
->max
) / 100;
517 this_dbs_info
->requested_freq
-= freq_target
;
518 if (this_dbs_info
->requested_freq
< policy
->min
)
519 this_dbs_info
->requested_freq
= policy
->min
;
522 * if we cannot reduce the frequency anymore, break out early
524 if (policy
->cur
== policy
->min
)
527 __cpufreq_driver_target(policy
, this_dbs_info
->requested_freq
,
533 static void do_dbs_timer(struct work_struct
*work
)
535 struct cpu_dbs_info_s
*dbs_info
=
536 container_of(work
, struct cpu_dbs_info_s
, work
.work
);
537 unsigned int cpu
= dbs_info
->cpu
;
539 /* We want all CPUs to do sampling nearly on same jiffy */
540 int delay
= usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
542 delay
-= jiffies
% delay
;
544 mutex_lock(&dbs_info
->timer_mutex
);
546 dbs_check_cpu(dbs_info
);
548 schedule_delayed_work_on(cpu
, &dbs_info
->work
, delay
);
549 mutex_unlock(&dbs_info
->timer_mutex
);
552 static inline void dbs_timer_init(struct cpu_dbs_info_s
*dbs_info
)
554 /* We want all CPUs to do sampling nearly on same jiffy */
555 int delay
= usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
556 delay
-= jiffies
% delay
;
558 dbs_info
->enable
= 1;
559 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info
->work
, do_dbs_timer
);
560 schedule_delayed_work_on(dbs_info
->cpu
, &dbs_info
->work
, delay
);
563 static inline void dbs_timer_exit(struct cpu_dbs_info_s
*dbs_info
)
565 dbs_info
->enable
= 0;
566 cancel_delayed_work_sync(&dbs_info
->work
);
569 static int cpufreq_governor_dbs(struct cpufreq_policy
*policy
,
572 unsigned int cpu
= policy
->cpu
;
573 struct cpu_dbs_info_s
*this_dbs_info
;
577 this_dbs_info
= &per_cpu(cs_cpu_dbs_info
, cpu
);
580 case CPUFREQ_GOV_START
:
581 if ((!cpu_online(cpu
)) || (!policy
->cur
))
584 mutex_lock(&dbs_mutex
);
586 rc
= sysfs_create_group(&policy
->kobj
, &dbs_attr_group_old
);
588 mutex_unlock(&dbs_mutex
);
592 for_each_cpu(j
, policy
->cpus
) {
593 struct cpu_dbs_info_s
*j_dbs_info
;
594 j_dbs_info
= &per_cpu(cs_cpu_dbs_info
, j
);
595 j_dbs_info
->cur_policy
= policy
;
597 j_dbs_info
->prev_cpu_idle
= get_cpu_idle_time(j
,
598 &j_dbs_info
->prev_cpu_wall
);
599 if (dbs_tuners_ins
.ignore_nice
) {
600 j_dbs_info
->prev_cpu_nice
=
601 kstat_cpu(j
).cpustat
.nice
;
604 this_dbs_info
->down_skip
= 0;
605 this_dbs_info
->requested_freq
= policy
->cur
;
607 mutex_init(&this_dbs_info
->timer_mutex
);
610 * Start the timerschedule work, when this governor
611 * is used for first time
613 if (dbs_enable
== 1) {
614 unsigned int latency
;
615 /* policy latency is in nS. Convert it to uS first */
616 latency
= policy
->cpuinfo
.transition_latency
/ 1000;
620 rc
= sysfs_create_group(cpufreq_global_kobject
,
623 mutex_unlock(&dbs_mutex
);
628 * conservative does not implement micro like ondemand
629 * governor, thus we are bound to jiffes/HZ
632 MIN_SAMPLING_RATE_RATIO
* jiffies_to_usecs(10);
633 /* Bring kernel and HW constraints together */
634 min_sampling_rate
= max(min_sampling_rate
,
635 MIN_LATENCY_MULTIPLIER
* latency
);
636 dbs_tuners_ins
.sampling_rate
=
637 max(min_sampling_rate
,
638 latency
* LATENCY_MULTIPLIER
);
640 cpufreq_register_notifier(
641 &dbs_cpufreq_notifier_block
,
642 CPUFREQ_TRANSITION_NOTIFIER
);
644 mutex_unlock(&dbs_mutex
);
646 dbs_timer_init(this_dbs_info
);
650 case CPUFREQ_GOV_STOP
:
651 dbs_timer_exit(this_dbs_info
);
653 mutex_lock(&dbs_mutex
);
654 sysfs_remove_group(&policy
->kobj
, &dbs_attr_group_old
);
656 mutex_destroy(&this_dbs_info
->timer_mutex
);
659 * Stop the timerschedule work, when this governor
660 * is used for first time
663 cpufreq_unregister_notifier(
664 &dbs_cpufreq_notifier_block
,
665 CPUFREQ_TRANSITION_NOTIFIER
);
667 mutex_unlock(&dbs_mutex
);
669 sysfs_remove_group(cpufreq_global_kobject
,
674 case CPUFREQ_GOV_LIMITS
:
675 mutex_lock(&this_dbs_info
->timer_mutex
);
676 if (policy
->max
< this_dbs_info
->cur_policy
->cur
)
677 __cpufreq_driver_target(
678 this_dbs_info
->cur_policy
,
679 policy
->max
, CPUFREQ_RELATION_H
);
680 else if (policy
->min
> this_dbs_info
->cur_policy
->cur
)
681 __cpufreq_driver_target(
682 this_dbs_info
->cur_policy
,
683 policy
->min
, CPUFREQ_RELATION_L
);
684 mutex_unlock(&this_dbs_info
->timer_mutex
);
691 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
694 struct cpufreq_governor cpufreq_gov_conservative
= {
695 .name
= "conservative",
696 .governor
= cpufreq_governor_dbs
,
697 .max_transition_latency
= TRANSITION_LATENCY_LIMIT
,
698 .owner
= THIS_MODULE
,
701 static int __init
cpufreq_gov_dbs_init(void)
703 return cpufreq_register_governor(&cpufreq_gov_conservative
);
706 static void __exit
cpufreq_gov_dbs_exit(void)
708 cpufreq_unregister_governor(&cpufreq_gov_conservative
);
712 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
713 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
714 "Low Latency Frequency Transition capable processors "
715 "optimised for use in a battery environment");
716 MODULE_LICENSE("GPL");
718 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
719 fs_initcall(cpufreq_gov_dbs_init
);
721 module_init(cpufreq_gov_dbs_init
);
723 module_exit(cpufreq_gov_dbs_exit
);