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/sched.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 with CPUFREQ_ETERNAL)
47 * 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 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
54 #define MIN_SAMPLING_RATE (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
55 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
56 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
57 #define DEF_SAMPLING_DOWN_FACTOR (1)
58 #define MAX_SAMPLING_DOWN_FACTOR (10)
59 #define TRANSITION_LATENCY_LIMIT (10 * 1000)
61 static void do_dbs_timer(void *data
);
63 struct cpu_dbs_info_s
{
64 struct cpufreq_policy
*cur_policy
;
65 unsigned int prev_cpu_idle_up
;
66 unsigned int prev_cpu_idle_down
;
68 unsigned int down_skip
;
69 unsigned int requested_freq
;
71 static DEFINE_PER_CPU(struct cpu_dbs_info_s
, cpu_dbs_info
);
73 static unsigned int dbs_enable
; /* number of CPUs using this policy */
75 static DEFINE_MUTEX (dbs_mutex
);
76 static DECLARE_WORK (dbs_work
, do_dbs_timer
, NULL
);
79 unsigned int sampling_rate
;
80 unsigned int sampling_down_factor
;
81 unsigned int up_threshold
;
82 unsigned int down_threshold
;
83 unsigned int ignore_nice
;
84 unsigned int freq_step
;
87 static struct dbs_tuners dbs_tuners_ins
= {
88 .up_threshold
= DEF_FREQUENCY_UP_THRESHOLD
,
89 .down_threshold
= DEF_FREQUENCY_DOWN_THRESHOLD
,
90 .sampling_down_factor
= DEF_SAMPLING_DOWN_FACTOR
,
95 static inline unsigned int get_cpu_idle_time(unsigned int cpu
)
97 return kstat_cpu(cpu
).cpustat
.idle
+
98 kstat_cpu(cpu
).cpustat
.iowait
+
99 ( dbs_tuners_ins
.ignore_nice
?
100 kstat_cpu(cpu
).cpustat
.nice
:
104 /************************** sysfs interface ************************/
105 static ssize_t
show_sampling_rate_max(struct cpufreq_policy
*policy
, char *buf
)
107 return sprintf (buf
, "%u\n", MAX_SAMPLING_RATE
);
110 static ssize_t
show_sampling_rate_min(struct cpufreq_policy
*policy
, char *buf
)
112 return sprintf (buf
, "%u\n", MIN_SAMPLING_RATE
);
115 #define define_one_ro(_name) \
116 static struct freq_attr _name = \
117 __ATTR(_name, 0444, show_##_name, NULL)
119 define_one_ro(sampling_rate_max
);
120 define_one_ro(sampling_rate_min
);
122 /* cpufreq_conservative Governor Tunables */
123 #define show_one(file_name, object) \
124 static ssize_t show_##file_name \
125 (struct cpufreq_policy *unused, char *buf) \
127 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
129 show_one(sampling_rate
, sampling_rate
);
130 show_one(sampling_down_factor
, sampling_down_factor
);
131 show_one(up_threshold
, up_threshold
);
132 show_one(down_threshold
, down_threshold
);
133 show_one(ignore_nice_load
, ignore_nice
);
134 show_one(freq_step
, freq_step
);
136 static ssize_t
store_sampling_down_factor(struct cpufreq_policy
*unused
,
137 const char *buf
, size_t count
)
141 ret
= sscanf (buf
, "%u", &input
);
142 if (ret
!= 1 || input
> MAX_SAMPLING_DOWN_FACTOR
|| input
< 1)
145 mutex_lock(&dbs_mutex
);
146 dbs_tuners_ins
.sampling_down_factor
= input
;
147 mutex_unlock(&dbs_mutex
);
152 static ssize_t
store_sampling_rate(struct cpufreq_policy
*unused
,
153 const char *buf
, size_t count
)
157 ret
= sscanf (buf
, "%u", &input
);
159 mutex_lock(&dbs_mutex
);
160 if (ret
!= 1 || input
> MAX_SAMPLING_RATE
|| input
< MIN_SAMPLING_RATE
) {
161 mutex_unlock(&dbs_mutex
);
165 dbs_tuners_ins
.sampling_rate
= input
;
166 mutex_unlock(&dbs_mutex
);
171 static ssize_t
store_up_threshold(struct cpufreq_policy
*unused
,
172 const char *buf
, size_t count
)
176 ret
= sscanf (buf
, "%u", &input
);
178 mutex_lock(&dbs_mutex
);
179 if (ret
!= 1 || input
> 100 || input
< 0 ||
180 input
<= dbs_tuners_ins
.down_threshold
) {
181 mutex_unlock(&dbs_mutex
);
185 dbs_tuners_ins
.up_threshold
= input
;
186 mutex_unlock(&dbs_mutex
);
191 static ssize_t
store_down_threshold(struct cpufreq_policy
*unused
,
192 const char *buf
, size_t count
)
196 ret
= sscanf (buf
, "%u", &input
);
198 mutex_lock(&dbs_mutex
);
199 if (ret
!= 1 || input
> 100 || input
< 0 ||
200 input
>= dbs_tuners_ins
.up_threshold
) {
201 mutex_unlock(&dbs_mutex
);
205 dbs_tuners_ins
.down_threshold
= input
;
206 mutex_unlock(&dbs_mutex
);
211 static ssize_t
store_ignore_nice_load(struct cpufreq_policy
*policy
,
212 const char *buf
, size_t count
)
219 ret
= sscanf (buf
, "%u", &input
);
226 mutex_lock(&dbs_mutex
);
227 if ( input
== dbs_tuners_ins
.ignore_nice
) { /* nothing to do */
228 mutex_unlock(&dbs_mutex
);
231 dbs_tuners_ins
.ignore_nice
= input
;
233 /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
234 for_each_online_cpu(j
) {
235 struct cpu_dbs_info_s
*j_dbs_info
;
236 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
237 j_dbs_info
->prev_cpu_idle_up
= get_cpu_idle_time(j
);
238 j_dbs_info
->prev_cpu_idle_down
= j_dbs_info
->prev_cpu_idle_up
;
240 mutex_unlock(&dbs_mutex
);
245 static ssize_t
store_freq_step(struct cpufreq_policy
*policy
,
246 const char *buf
, size_t count
)
251 ret
= sscanf (buf
, "%u", &input
);
259 /* no need to test here if freq_step is zero as the user might actually
260 * want this, they would be crazy though :) */
261 mutex_lock(&dbs_mutex
);
262 dbs_tuners_ins
.freq_step
= input
;
263 mutex_unlock(&dbs_mutex
);
268 #define define_one_rw(_name) \
269 static struct freq_attr _name = \
270 __ATTR(_name, 0644, show_##_name, store_##_name)
272 define_one_rw(sampling_rate
);
273 define_one_rw(sampling_down_factor
);
274 define_one_rw(up_threshold
);
275 define_one_rw(down_threshold
);
276 define_one_rw(ignore_nice_load
);
277 define_one_rw(freq_step
);
279 static struct attribute
* dbs_attributes
[] = {
280 &sampling_rate_max
.attr
,
281 &sampling_rate_min
.attr
,
283 &sampling_down_factor
.attr
,
285 &down_threshold
.attr
,
286 &ignore_nice_load
.attr
,
291 static struct attribute_group dbs_attr_group
= {
292 .attrs
= dbs_attributes
,
293 .name
= "conservative",
296 /************************** sysfs end ************************/
298 static void dbs_check_cpu(int cpu
)
300 unsigned int idle_ticks
, up_idle_ticks
, down_idle_ticks
;
301 unsigned int tmp_idle_ticks
, total_idle_ticks
;
302 unsigned int freq_step
;
303 unsigned int freq_down_sampling_rate
;
304 struct cpu_dbs_info_s
*this_dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
305 struct cpufreq_policy
*policy
;
307 if (!this_dbs_info
->enable
)
310 policy
= this_dbs_info
->cur_policy
;
313 * The default safe range is 20% to 80%
314 * Every sampling_rate, we check
315 * - If current idle time is less than 20%, then we try to
317 * Every sampling_rate*sampling_down_factor, we check
318 * - If current idle time is more than 80%, then we try to
321 * Any frequency increase takes it to the maximum frequency.
322 * Frequency reduction happens at minimum steps of
323 * 5% (default) of max_frequency
326 /* Check for frequency increase */
327 idle_ticks
= UINT_MAX
;
329 /* Check for frequency increase */
330 total_idle_ticks
= get_cpu_idle_time(cpu
);
331 tmp_idle_ticks
= total_idle_ticks
-
332 this_dbs_info
->prev_cpu_idle_up
;
333 this_dbs_info
->prev_cpu_idle_up
= total_idle_ticks
;
335 if (tmp_idle_ticks
< idle_ticks
)
336 idle_ticks
= tmp_idle_ticks
;
338 /* Scale idle ticks by 100 and compare with up and down ticks */
340 up_idle_ticks
= (100 - dbs_tuners_ins
.up_threshold
) *
341 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
343 if (idle_ticks
< up_idle_ticks
) {
344 this_dbs_info
->down_skip
= 0;
345 this_dbs_info
->prev_cpu_idle_down
=
346 this_dbs_info
->prev_cpu_idle_up
;
348 /* if we are already at full speed then break out early */
349 if (this_dbs_info
->requested_freq
== policy
->max
)
352 freq_step
= (dbs_tuners_ins
.freq_step
* policy
->max
) / 100;
354 /* max freq cannot be less than 100. But who knows.... */
355 if (unlikely(freq_step
== 0))
358 this_dbs_info
->requested_freq
+= freq_step
;
359 if (this_dbs_info
->requested_freq
> policy
->max
)
360 this_dbs_info
->requested_freq
= policy
->max
;
362 __cpufreq_driver_target(policy
, this_dbs_info
->requested_freq
,
367 /* Check for frequency decrease */
368 this_dbs_info
->down_skip
++;
369 if (this_dbs_info
->down_skip
< dbs_tuners_ins
.sampling_down_factor
)
372 /* Check for frequency decrease */
373 total_idle_ticks
= this_dbs_info
->prev_cpu_idle_up
;
374 tmp_idle_ticks
= total_idle_ticks
-
375 this_dbs_info
->prev_cpu_idle_down
;
376 this_dbs_info
->prev_cpu_idle_down
= total_idle_ticks
;
378 if (tmp_idle_ticks
< idle_ticks
)
379 idle_ticks
= tmp_idle_ticks
;
381 /* Scale idle ticks by 100 and compare with up and down ticks */
383 this_dbs_info
->down_skip
= 0;
385 freq_down_sampling_rate
= dbs_tuners_ins
.sampling_rate
*
386 dbs_tuners_ins
.sampling_down_factor
;
387 down_idle_ticks
= (100 - dbs_tuners_ins
.down_threshold
) *
388 usecs_to_jiffies(freq_down_sampling_rate
);
390 if (idle_ticks
> down_idle_ticks
) {
392 * if we are already at the lowest speed then break out early
393 * or if we 'cannot' reduce the speed as the user might want
394 * freq_step to be zero
396 if (this_dbs_info
->requested_freq
== policy
->min
397 || dbs_tuners_ins
.freq_step
== 0)
400 freq_step
= (dbs_tuners_ins
.freq_step
* policy
->max
) / 100;
402 /* max freq cannot be less than 100. But who knows.... */
403 if (unlikely(freq_step
== 0))
406 this_dbs_info
->requested_freq
-= freq_step
;
407 if (this_dbs_info
->requested_freq
< policy
->min
)
408 this_dbs_info
->requested_freq
= policy
->min
;
410 __cpufreq_driver_target(policy
, this_dbs_info
->requested_freq
,
416 static void do_dbs_timer(void *data
)
419 mutex_lock(&dbs_mutex
);
420 for_each_online_cpu(i
)
422 schedule_delayed_work(&dbs_work
,
423 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
));
424 mutex_unlock(&dbs_mutex
);
427 static inline void dbs_timer_init(void)
429 INIT_WORK(&dbs_work
, do_dbs_timer
, NULL
);
430 schedule_delayed_work(&dbs_work
,
431 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
));
435 static inline void dbs_timer_exit(void)
437 cancel_delayed_work(&dbs_work
);
441 static int cpufreq_governor_dbs(struct cpufreq_policy
*policy
,
444 unsigned int cpu
= policy
->cpu
;
445 struct cpu_dbs_info_s
*this_dbs_info
;
448 this_dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
451 case CPUFREQ_GOV_START
:
452 if ((!cpu_online(cpu
)) ||
456 if (policy
->cpuinfo
.transition_latency
>
457 (TRANSITION_LATENCY_LIMIT
* 1000))
459 if (this_dbs_info
->enable
) /* Already enabled */
462 mutex_lock(&dbs_mutex
);
463 for_each_cpu_mask(j
, policy
->cpus
) {
464 struct cpu_dbs_info_s
*j_dbs_info
;
465 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
466 j_dbs_info
->cur_policy
= policy
;
468 j_dbs_info
->prev_cpu_idle_up
= get_cpu_idle_time(cpu
);
469 j_dbs_info
->prev_cpu_idle_down
470 = j_dbs_info
->prev_cpu_idle_up
;
472 this_dbs_info
->enable
= 1;
473 this_dbs_info
->down_skip
= 0;
474 this_dbs_info
->requested_freq
= policy
->cur
;
475 sysfs_create_group(&policy
->kobj
, &dbs_attr_group
);
478 * Start the timerschedule work, when this governor
479 * is used for first time
481 if (dbs_enable
== 1) {
482 unsigned int latency
;
483 /* policy latency is in nS. Convert it to uS first */
484 latency
= policy
->cpuinfo
.transition_latency
/ 1000;
488 def_sampling_rate
= 10 * latency
*
489 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER
;
491 if (def_sampling_rate
< MIN_STAT_SAMPLING_RATE
)
492 def_sampling_rate
= MIN_STAT_SAMPLING_RATE
;
494 dbs_tuners_ins
.sampling_rate
= def_sampling_rate
;
499 mutex_unlock(&dbs_mutex
);
502 case CPUFREQ_GOV_STOP
:
503 mutex_lock(&dbs_mutex
);
504 this_dbs_info
->enable
= 0;
505 sysfs_remove_group(&policy
->kobj
, &dbs_attr_group
);
508 * Stop the timerschedule work, when this governor
509 * is used for first time
514 mutex_unlock(&dbs_mutex
);
518 case CPUFREQ_GOV_LIMITS
:
519 mutex_lock(&dbs_mutex
);
520 if (policy
->max
< this_dbs_info
->cur_policy
->cur
)
521 __cpufreq_driver_target(
522 this_dbs_info
->cur_policy
,
523 policy
->max
, CPUFREQ_RELATION_H
);
524 else if (policy
->min
> this_dbs_info
->cur_policy
->cur
)
525 __cpufreq_driver_target(
526 this_dbs_info
->cur_policy
,
527 policy
->min
, CPUFREQ_RELATION_L
);
528 mutex_unlock(&dbs_mutex
);
534 static struct cpufreq_governor cpufreq_gov_dbs
= {
535 .name
= "conservative",
536 .governor
= cpufreq_governor_dbs
,
537 .owner
= THIS_MODULE
,
540 static int __init
cpufreq_gov_dbs_init(void)
542 return cpufreq_register_governor(&cpufreq_gov_dbs
);
545 static void __exit
cpufreq_gov_dbs_exit(void)
547 /* Make sure that the scheduled work is indeed not running */
548 flush_scheduled_work();
550 cpufreq_unregister_governor(&cpufreq_gov_dbs
);
554 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
555 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
556 "Low Latency Frequency Transition capable processors "
557 "optimised for use in a battery environment");
558 MODULE_LICENSE ("GPL");
560 module_init(cpufreq_gov_dbs_init
);
561 module_exit(cpufreq_gov_dbs_exit
);