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) 2004 Alexander Clouter <alex-kernel@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/smp.h>
17 #include <linux/init.h>
18 #include <linux/interrupt.h>
19 #include <linux/ctype.h>
20 #include <linux/cpufreq.h>
21 #include <linux/sysctl.h>
22 #include <linux/types.h>
24 #include <linux/sysfs.h>
25 #include <linux/cpu.h>
26 #include <linux/kmod.h>
27 #include <linux/workqueue.h>
28 #include <linux/jiffies.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/percpu.h>
31 #include <linux/mutex.h>
33 * dbs is used in this file as a shortform for demandbased switching
34 * It helps to keep variable names smaller, simpler
37 #define DEF_FREQUENCY_UP_THRESHOLD (80)
38 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
41 * The polling frequency of this governor depends on the capability of
42 * the processor. Default polling frequency is 1000 times the transition
43 * latency of the processor. The governor will work on any processor with
44 * transition latency <= 10mS, using appropriate sampling
46 * For CPUs with transition latency > 10mS (mostly drivers
47 * with CPUFREQ_ETERNAL), this governor will not work.
48 * All times here are in uS.
50 static unsigned int def_sampling_rate
;
51 #define MIN_SAMPLING_RATE_RATIO (2)
52 /* for correct statistics, we need at least 10 ticks between each measure */
53 #define MIN_STAT_SAMPLING_RATE \
54 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
55 #define MIN_SAMPLING_RATE \
56 (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
57 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
58 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
59 #define DEF_SAMPLING_DOWN_FACTOR (1)
60 #define MAX_SAMPLING_DOWN_FACTOR (10)
61 #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
63 static void do_dbs_timer(struct work_struct
*work
);
65 struct cpu_dbs_info_s
{
66 struct cpufreq_policy
*cur_policy
;
67 unsigned int prev_cpu_idle_up
;
68 unsigned int prev_cpu_idle_down
;
70 unsigned int down_skip
;
71 unsigned int requested_freq
;
73 static DEFINE_PER_CPU(struct cpu_dbs_info_s
, cpu_dbs_info
);
75 static unsigned int dbs_enable
; /* number of CPUs using this policy */
78 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
79 * lock and dbs_mutex. cpu_hotplug lock should always be held before
80 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
81 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
82 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
83 * is recursive for the same process. -Venki
85 static DEFINE_MUTEX(dbs_mutex
);
86 static DECLARE_DELAYED_WORK(dbs_work
, do_dbs_timer
);
89 unsigned int sampling_rate
;
90 unsigned int sampling_down_factor
;
91 unsigned int up_threshold
;
92 unsigned int down_threshold
;
93 unsigned int ignore_nice
;
94 unsigned int freq_step
;
97 static struct dbs_tuners dbs_tuners_ins
= {
98 .up_threshold
= DEF_FREQUENCY_UP_THRESHOLD
,
99 .down_threshold
= DEF_FREQUENCY_DOWN_THRESHOLD
,
100 .sampling_down_factor
= DEF_SAMPLING_DOWN_FACTOR
,
105 static inline unsigned int get_cpu_idle_time(unsigned int cpu
)
107 unsigned int add_nice
= 0, ret
;
109 if (dbs_tuners_ins
.ignore_nice
)
110 add_nice
= kstat_cpu(cpu
).cpustat
.nice
;
112 ret
= kstat_cpu(cpu
).cpustat
.idle
+
113 kstat_cpu(cpu
).cpustat
.iowait
+
119 /* keep track of frequency transitions */
121 dbs_cpufreq_notifier(struct notifier_block
*nb
, unsigned long val
,
124 struct cpufreq_freqs
*freq
= data
;
125 struct cpu_dbs_info_s
*this_dbs_info
= &per_cpu(cpu_dbs_info
,
128 if (!this_dbs_info
->enable
)
131 this_dbs_info
->requested_freq
= freq
->new;
136 static struct notifier_block dbs_cpufreq_notifier_block
= {
137 .notifier_call
= dbs_cpufreq_notifier
140 /************************** sysfs interface ************************/
141 static ssize_t
show_sampling_rate_max(struct cpufreq_policy
*policy
, char *buf
)
143 return sprintf(buf
, "%u\n", MAX_SAMPLING_RATE
);
146 static ssize_t
show_sampling_rate_min(struct cpufreq_policy
*policy
, char *buf
)
148 return sprintf(buf
, "%u\n", MIN_SAMPLING_RATE
);
151 #define define_one_ro(_name) \
152 static struct freq_attr _name = \
153 __ATTR(_name, 0444, show_##_name, NULL)
155 define_one_ro(sampling_rate_max
);
156 define_one_ro(sampling_rate_min
);
158 /* cpufreq_conservative Governor Tunables */
159 #define show_one(file_name, object) \
160 static ssize_t show_##file_name \
161 (struct cpufreq_policy *unused, char *buf) \
163 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
165 show_one(sampling_rate
, sampling_rate
);
166 show_one(sampling_down_factor
, sampling_down_factor
);
167 show_one(up_threshold
, up_threshold
);
168 show_one(down_threshold
, down_threshold
);
169 show_one(ignore_nice_load
, ignore_nice
);
170 show_one(freq_step
, freq_step
);
172 static ssize_t
store_sampling_down_factor(struct cpufreq_policy
*unused
,
173 const char *buf
, size_t count
)
177 ret
= sscanf(buf
, "%u", &input
);
178 if (ret
!= 1 || input
> MAX_SAMPLING_DOWN_FACTOR
|| input
< 1)
181 mutex_lock(&dbs_mutex
);
182 dbs_tuners_ins
.sampling_down_factor
= input
;
183 mutex_unlock(&dbs_mutex
);
188 static ssize_t
store_sampling_rate(struct cpufreq_policy
*unused
,
189 const char *buf
, size_t count
)
193 ret
= sscanf(buf
, "%u", &input
);
195 mutex_lock(&dbs_mutex
);
196 if (ret
!= 1 || input
> MAX_SAMPLING_RATE
||
197 input
< MIN_SAMPLING_RATE
) {
198 mutex_unlock(&dbs_mutex
);
202 dbs_tuners_ins
.sampling_rate
= input
;
203 mutex_unlock(&dbs_mutex
);
208 static ssize_t
store_up_threshold(struct cpufreq_policy
*unused
,
209 const char *buf
, size_t count
)
213 ret
= sscanf(buf
, "%u", &input
);
215 mutex_lock(&dbs_mutex
);
216 if (ret
!= 1 || input
> 100 ||
217 input
<= dbs_tuners_ins
.down_threshold
) {
218 mutex_unlock(&dbs_mutex
);
222 dbs_tuners_ins
.up_threshold
= input
;
223 mutex_unlock(&dbs_mutex
);
228 static ssize_t
store_down_threshold(struct cpufreq_policy
*unused
,
229 const char *buf
, size_t count
)
233 ret
= sscanf(buf
, "%u", &input
);
235 mutex_lock(&dbs_mutex
);
236 if (ret
!= 1 || input
> 100 || input
>= dbs_tuners_ins
.up_threshold
) {
237 mutex_unlock(&dbs_mutex
);
241 dbs_tuners_ins
.down_threshold
= input
;
242 mutex_unlock(&dbs_mutex
);
247 static ssize_t
store_ignore_nice_load(struct cpufreq_policy
*policy
,
248 const char *buf
, size_t count
)
255 ret
= sscanf(buf
, "%u", &input
);
262 mutex_lock(&dbs_mutex
);
263 if (input
== dbs_tuners_ins
.ignore_nice
) { /* nothing to do */
264 mutex_unlock(&dbs_mutex
);
267 dbs_tuners_ins
.ignore_nice
= input
;
269 /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
270 for_each_online_cpu(j
) {
271 struct cpu_dbs_info_s
*j_dbs_info
;
272 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
273 j_dbs_info
->prev_cpu_idle_up
= get_cpu_idle_time(j
);
274 j_dbs_info
->prev_cpu_idle_down
= j_dbs_info
->prev_cpu_idle_up
;
276 mutex_unlock(&dbs_mutex
);
281 static ssize_t
store_freq_step(struct cpufreq_policy
*policy
,
282 const char *buf
, size_t count
)
287 ret
= sscanf(buf
, "%u", &input
);
295 /* no need to test here if freq_step is zero as the user might actually
296 * want this, they would be crazy though :) */
297 mutex_lock(&dbs_mutex
);
298 dbs_tuners_ins
.freq_step
= input
;
299 mutex_unlock(&dbs_mutex
);
304 #define define_one_rw(_name) \
305 static struct freq_attr _name = \
306 __ATTR(_name, 0644, show_##_name, store_##_name)
308 define_one_rw(sampling_rate
);
309 define_one_rw(sampling_down_factor
);
310 define_one_rw(up_threshold
);
311 define_one_rw(down_threshold
);
312 define_one_rw(ignore_nice_load
);
313 define_one_rw(freq_step
);
315 static struct attribute
*dbs_attributes
[] = {
316 &sampling_rate_max
.attr
,
317 &sampling_rate_min
.attr
,
319 &sampling_down_factor
.attr
,
321 &down_threshold
.attr
,
322 &ignore_nice_load
.attr
,
327 static struct attribute_group dbs_attr_group
= {
328 .attrs
= dbs_attributes
,
329 .name
= "conservative",
332 /************************** sysfs end ************************/
334 static void dbs_check_cpu(int cpu
)
336 unsigned int idle_ticks
, up_idle_ticks
, down_idle_ticks
;
337 unsigned int tmp_idle_ticks
, total_idle_ticks
;
338 unsigned int freq_target
;
339 unsigned int freq_down_sampling_rate
;
340 struct cpu_dbs_info_s
*this_dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
341 struct cpufreq_policy
*policy
;
343 if (!this_dbs_info
->enable
)
346 policy
= this_dbs_info
->cur_policy
;
349 * The default safe range is 20% to 80%
350 * Every sampling_rate, we check
351 * - If current idle time is less than 20%, then we try to
353 * Every sampling_rate*sampling_down_factor, we check
354 * - If current idle time is more than 80%, then we try to
357 * Any frequency increase takes it to the maximum frequency.
358 * Frequency reduction happens at minimum steps of
359 * 5% (default) of max_frequency
362 /* Check for frequency increase */
363 idle_ticks
= UINT_MAX
;
365 /* Check for frequency increase */
366 total_idle_ticks
= get_cpu_idle_time(cpu
);
367 tmp_idle_ticks
= total_idle_ticks
-
368 this_dbs_info
->prev_cpu_idle_up
;
369 this_dbs_info
->prev_cpu_idle_up
= total_idle_ticks
;
371 if (tmp_idle_ticks
< idle_ticks
)
372 idle_ticks
= tmp_idle_ticks
;
374 /* Scale idle ticks by 100 and compare with up and down ticks */
376 up_idle_ticks
= (100 - dbs_tuners_ins
.up_threshold
) *
377 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
379 if (idle_ticks
< up_idle_ticks
) {
380 this_dbs_info
->down_skip
= 0;
381 this_dbs_info
->prev_cpu_idle_down
=
382 this_dbs_info
->prev_cpu_idle_up
;
384 /* if we are already at full speed then break out early */
385 if (this_dbs_info
->requested_freq
== policy
->max
)
388 freq_target
= (dbs_tuners_ins
.freq_step
* policy
->max
) / 100;
390 /* max freq cannot be less than 100. But who knows.... */
391 if (unlikely(freq_target
== 0))
394 this_dbs_info
->requested_freq
+= freq_target
;
395 if (this_dbs_info
->requested_freq
> policy
->max
)
396 this_dbs_info
->requested_freq
= policy
->max
;
398 __cpufreq_driver_target(policy
, this_dbs_info
->requested_freq
,
403 /* Check for frequency decrease */
404 this_dbs_info
->down_skip
++;
405 if (this_dbs_info
->down_skip
< dbs_tuners_ins
.sampling_down_factor
)
408 /* Check for frequency decrease */
409 total_idle_ticks
= this_dbs_info
->prev_cpu_idle_up
;
410 tmp_idle_ticks
= total_idle_ticks
-
411 this_dbs_info
->prev_cpu_idle_down
;
412 this_dbs_info
->prev_cpu_idle_down
= total_idle_ticks
;
414 if (tmp_idle_ticks
< idle_ticks
)
415 idle_ticks
= tmp_idle_ticks
;
417 /* Scale idle ticks by 100 and compare with up and down ticks */
419 this_dbs_info
->down_skip
= 0;
421 freq_down_sampling_rate
= dbs_tuners_ins
.sampling_rate
*
422 dbs_tuners_ins
.sampling_down_factor
;
423 down_idle_ticks
= (100 - dbs_tuners_ins
.down_threshold
) *
424 usecs_to_jiffies(freq_down_sampling_rate
);
426 if (idle_ticks
> down_idle_ticks
) {
428 * if we are already at the lowest speed then break out early
429 * or if we 'cannot' reduce the speed as the user might want
430 * freq_target to be zero
432 if (this_dbs_info
->requested_freq
== policy
->min
433 || dbs_tuners_ins
.freq_step
== 0)
436 freq_target
= (dbs_tuners_ins
.freq_step
* policy
->max
) / 100;
438 /* max freq cannot be less than 100. But who knows.... */
439 if (unlikely(freq_target
== 0))
442 this_dbs_info
->requested_freq
-= freq_target
;
443 if (this_dbs_info
->requested_freq
< policy
->min
)
444 this_dbs_info
->requested_freq
= policy
->min
;
446 __cpufreq_driver_target(policy
, this_dbs_info
->requested_freq
,
452 static void do_dbs_timer(struct work_struct
*work
)
455 mutex_lock(&dbs_mutex
);
456 for_each_online_cpu(i
)
458 schedule_delayed_work(&dbs_work
,
459 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
));
460 mutex_unlock(&dbs_mutex
);
463 static inline void dbs_timer_init(void)
465 init_timer_deferrable(&dbs_work
.timer
);
466 schedule_delayed_work(&dbs_work
,
467 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
));
471 static inline void dbs_timer_exit(void)
473 cancel_delayed_work(&dbs_work
);
477 static int cpufreq_governor_dbs(struct cpufreq_policy
*policy
,
480 unsigned int cpu
= policy
->cpu
;
481 struct cpu_dbs_info_s
*this_dbs_info
;
485 this_dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
488 case CPUFREQ_GOV_START
:
489 if ((!cpu_online(cpu
)) || (!policy
->cur
))
492 if (this_dbs_info
->enable
) /* Already enabled */
495 mutex_lock(&dbs_mutex
);
497 rc
= sysfs_create_group(&policy
->kobj
, &dbs_attr_group
);
499 mutex_unlock(&dbs_mutex
);
503 for_each_cpu(j
, policy
->cpus
) {
504 struct cpu_dbs_info_s
*j_dbs_info
;
505 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
506 j_dbs_info
->cur_policy
= policy
;
508 j_dbs_info
->prev_cpu_idle_up
= get_cpu_idle_time(cpu
);
509 j_dbs_info
->prev_cpu_idle_down
510 = j_dbs_info
->prev_cpu_idle_up
;
512 this_dbs_info
->enable
= 1;
513 this_dbs_info
->down_skip
= 0;
514 this_dbs_info
->requested_freq
= policy
->cur
;
518 * Start the timerschedule work, when this governor
519 * is used for first time
521 if (dbs_enable
== 1) {
522 unsigned int latency
;
523 /* policy latency is in nS. Convert it to uS first */
524 latency
= policy
->cpuinfo
.transition_latency
/ 1000;
528 def_sampling_rate
= 10 * latency
*
529 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER
;
531 if (def_sampling_rate
< MIN_STAT_SAMPLING_RATE
)
532 def_sampling_rate
= MIN_STAT_SAMPLING_RATE
;
534 dbs_tuners_ins
.sampling_rate
= def_sampling_rate
;
537 cpufreq_register_notifier(
538 &dbs_cpufreq_notifier_block
,
539 CPUFREQ_TRANSITION_NOTIFIER
);
542 mutex_unlock(&dbs_mutex
);
545 case CPUFREQ_GOV_STOP
:
546 mutex_lock(&dbs_mutex
);
547 this_dbs_info
->enable
= 0;
548 sysfs_remove_group(&policy
->kobj
, &dbs_attr_group
);
551 * Stop the timerschedule work, when this governor
552 * is used for first time
554 if (dbs_enable
== 0) {
556 cpufreq_unregister_notifier(
557 &dbs_cpufreq_notifier_block
,
558 CPUFREQ_TRANSITION_NOTIFIER
);
561 mutex_unlock(&dbs_mutex
);
565 case CPUFREQ_GOV_LIMITS
:
566 mutex_lock(&dbs_mutex
);
567 if (policy
->max
< this_dbs_info
->cur_policy
->cur
)
568 __cpufreq_driver_target(
569 this_dbs_info
->cur_policy
,
570 policy
->max
, CPUFREQ_RELATION_H
);
571 else if (policy
->min
> this_dbs_info
->cur_policy
->cur
)
572 __cpufreq_driver_target(
573 this_dbs_info
->cur_policy
,
574 policy
->min
, CPUFREQ_RELATION_L
);
575 mutex_unlock(&dbs_mutex
);
581 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
584 struct cpufreq_governor cpufreq_gov_conservative
= {
585 .name
= "conservative",
586 .governor
= cpufreq_governor_dbs
,
587 .max_transition_latency
= TRANSITION_LATENCY_LIMIT
,
588 .owner
= THIS_MODULE
,
591 static int __init
cpufreq_gov_dbs_init(void)
593 return cpufreq_register_governor(&cpufreq_gov_conservative
);
596 static void __exit
cpufreq_gov_dbs_exit(void)
598 /* Make sure that the scheduled work is indeed not running */
599 flush_scheduled_work();
601 cpufreq_unregister_governor(&cpufreq_gov_conservative
);
605 MODULE_AUTHOR("Alexander Clouter <alex-kernel@digriz.org.uk>");
606 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
607 "Low Latency Frequency Transition capable processors "
608 "optimised for use in a battery environment");
609 MODULE_LICENSE("GPL");
611 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
612 fs_initcall(cpufreq_gov_dbs_init
);
614 module_init(cpufreq_gov_dbs_init
);
616 module_exit(cpufreq_gov_dbs_exit
);