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 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/init.h>
16 #include <linux/cpufreq.h>
17 #include <linux/cpu.h>
18 #include <linux/jiffies.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mutex.h>
23 * dbs is used in this file as a shortform for demandbased switching
24 * It helps to keep variable names smaller, simpler
27 #define DEF_FREQUENCY_UP_THRESHOLD (80)
28 #define MIN_FREQUENCY_UP_THRESHOLD (11)
29 #define MAX_FREQUENCY_UP_THRESHOLD (100)
32 * The polling frequency of this governor depends on the capability of
33 * the processor. Default polling frequency is 1000 times the transition
34 * latency of the processor. The governor will work on any processor with
35 * transition latency <= 10mS, using appropriate sampling
37 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
38 * this governor will not work.
39 * All times here are in uS.
41 static unsigned int def_sampling_rate
;
42 #define MIN_SAMPLING_RATE_RATIO (2)
43 /* for correct statistics, we need at least 10 ticks between each measure */
44 #define MIN_STAT_SAMPLING_RATE \
45 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
46 #define MIN_SAMPLING_RATE \
47 (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
48 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
49 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
50 #define TRANSITION_LATENCY_LIMIT (10 * 1000)
52 static void do_dbs_timer(struct work_struct
*work
);
55 enum dbs_sample
{DBS_NORMAL_SAMPLE
, DBS_SUB_SAMPLE
};
57 struct cpu_dbs_info_s
{
58 cputime64_t prev_cpu_idle
;
59 cputime64_t prev_cpu_wall
;
60 struct cpufreq_policy
*cur_policy
;
61 struct delayed_work work
;
62 enum dbs_sample sample_type
;
64 struct cpufreq_frequency_table
*freq_table
;
66 unsigned int freq_lo_jiffies
;
67 unsigned int freq_hi_jiffies
;
69 static DEFINE_PER_CPU(struct cpu_dbs_info_s
, cpu_dbs_info
);
71 static unsigned int dbs_enable
; /* number of CPUs using this policy */
74 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
75 * lock and dbs_mutex. cpu_hotplug lock should always be held before
76 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
77 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
78 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
79 * is recursive for the same process. -Venki
81 static DEFINE_MUTEX(dbs_mutex
);
83 static struct workqueue_struct
*kondemand_wq
;
85 static struct dbs_tuners
{
86 unsigned int sampling_rate
;
87 unsigned int up_threshold
;
88 unsigned int ignore_nice
;
89 unsigned int powersave_bias
;
91 .up_threshold
= DEF_FREQUENCY_UP_THRESHOLD
,
96 static inline cputime64_t
get_cpu_idle_time(unsigned int cpu
)
100 retval
= cputime64_add(kstat_cpu(cpu
).cpustat
.idle
,
101 kstat_cpu(cpu
).cpustat
.iowait
);
103 if (dbs_tuners_ins
.ignore_nice
)
104 retval
= cputime64_add(retval
, kstat_cpu(cpu
).cpustat
.nice
);
110 * Find right freq to be set now with powersave_bias on.
111 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
112 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
114 static unsigned int powersave_bias_target(struct cpufreq_policy
*policy
,
115 unsigned int freq_next
,
116 unsigned int relation
)
118 unsigned int freq_req
, freq_reduc
, freq_avg
;
119 unsigned int freq_hi
, freq_lo
;
120 unsigned int index
= 0;
121 unsigned int jiffies_total
, jiffies_hi
, jiffies_lo
;
122 struct cpu_dbs_info_s
*dbs_info
= &per_cpu(cpu_dbs_info
, policy
->cpu
);
124 if (!dbs_info
->freq_table
) {
125 dbs_info
->freq_lo
= 0;
126 dbs_info
->freq_lo_jiffies
= 0;
130 cpufreq_frequency_table_target(policy
, dbs_info
->freq_table
, freq_next
,
132 freq_req
= dbs_info
->freq_table
[index
].frequency
;
133 freq_reduc
= freq_req
* dbs_tuners_ins
.powersave_bias
/ 1000;
134 freq_avg
= freq_req
- freq_reduc
;
136 /* Find freq bounds for freq_avg in freq_table */
138 cpufreq_frequency_table_target(policy
, dbs_info
->freq_table
, freq_avg
,
139 CPUFREQ_RELATION_H
, &index
);
140 freq_lo
= dbs_info
->freq_table
[index
].frequency
;
142 cpufreq_frequency_table_target(policy
, dbs_info
->freq_table
, freq_avg
,
143 CPUFREQ_RELATION_L
, &index
);
144 freq_hi
= dbs_info
->freq_table
[index
].frequency
;
146 /* Find out how long we have to be in hi and lo freqs */
147 if (freq_hi
== freq_lo
) {
148 dbs_info
->freq_lo
= 0;
149 dbs_info
->freq_lo_jiffies
= 0;
152 jiffies_total
= usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
153 jiffies_hi
= (freq_avg
- freq_lo
) * jiffies_total
;
154 jiffies_hi
+= ((freq_hi
- freq_lo
) / 2);
155 jiffies_hi
/= (freq_hi
- freq_lo
);
156 jiffies_lo
= jiffies_total
- jiffies_hi
;
157 dbs_info
->freq_lo
= freq_lo
;
158 dbs_info
->freq_lo_jiffies
= jiffies_lo
;
159 dbs_info
->freq_hi_jiffies
= jiffies_hi
;
163 static void ondemand_powersave_bias_init(void)
166 for_each_online_cpu(i
) {
167 struct cpu_dbs_info_s
*dbs_info
= &per_cpu(cpu_dbs_info
, i
);
168 dbs_info
->freq_table
= cpufreq_frequency_get_table(i
);
169 dbs_info
->freq_lo
= 0;
173 /************************** sysfs interface ************************/
174 static ssize_t
show_sampling_rate_max(struct cpufreq_policy
*policy
, char *buf
)
176 return sprintf (buf
, "%u\n", MAX_SAMPLING_RATE
);
179 static ssize_t
show_sampling_rate_min(struct cpufreq_policy
*policy
, char *buf
)
181 return sprintf (buf
, "%u\n", MIN_SAMPLING_RATE
);
184 #define define_one_ro(_name) \
185 static struct freq_attr _name = \
186 __ATTR(_name, 0444, show_##_name, NULL)
188 define_one_ro(sampling_rate_max
);
189 define_one_ro(sampling_rate_min
);
191 /* cpufreq_ondemand Governor Tunables */
192 #define show_one(file_name, object) \
193 static ssize_t show_##file_name \
194 (struct cpufreq_policy *unused, char *buf) \
196 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
198 show_one(sampling_rate
, sampling_rate
);
199 show_one(up_threshold
, up_threshold
);
200 show_one(ignore_nice_load
, ignore_nice
);
201 show_one(powersave_bias
, powersave_bias
);
203 static ssize_t
store_sampling_rate(struct cpufreq_policy
*unused
,
204 const char *buf
, size_t count
)
208 ret
= sscanf(buf
, "%u", &input
);
210 mutex_lock(&dbs_mutex
);
211 if (ret
!= 1 || input
> MAX_SAMPLING_RATE
212 || input
< MIN_SAMPLING_RATE
) {
213 mutex_unlock(&dbs_mutex
);
217 dbs_tuners_ins
.sampling_rate
= input
;
218 mutex_unlock(&dbs_mutex
);
223 static ssize_t
store_up_threshold(struct cpufreq_policy
*unused
,
224 const char *buf
, size_t count
)
228 ret
= sscanf(buf
, "%u", &input
);
230 mutex_lock(&dbs_mutex
);
231 if (ret
!= 1 || input
> MAX_FREQUENCY_UP_THRESHOLD
||
232 input
< MIN_FREQUENCY_UP_THRESHOLD
) {
233 mutex_unlock(&dbs_mutex
);
237 dbs_tuners_ins
.up_threshold
= input
;
238 mutex_unlock(&dbs_mutex
);
243 static ssize_t
store_ignore_nice_load(struct cpufreq_policy
*policy
,
244 const char *buf
, size_t count
)
251 ret
= sscanf(buf
, "%u", &input
);
258 mutex_lock(&dbs_mutex
);
259 if ( input
== dbs_tuners_ins
.ignore_nice
) { /* nothing to do */
260 mutex_unlock(&dbs_mutex
);
263 dbs_tuners_ins
.ignore_nice
= input
;
265 /* we need to re-evaluate prev_cpu_idle */
266 for_each_online_cpu(j
) {
267 struct cpu_dbs_info_s
*dbs_info
;
268 dbs_info
= &per_cpu(cpu_dbs_info
, j
);
269 dbs_info
->prev_cpu_idle
= get_cpu_idle_time(j
);
270 dbs_info
->prev_cpu_wall
= get_jiffies_64();
272 mutex_unlock(&dbs_mutex
);
277 static ssize_t
store_powersave_bias(struct cpufreq_policy
*unused
,
278 const char *buf
, size_t count
)
282 ret
= sscanf(buf
, "%u", &input
);
290 mutex_lock(&dbs_mutex
);
291 dbs_tuners_ins
.powersave_bias
= input
;
292 ondemand_powersave_bias_init();
293 mutex_unlock(&dbs_mutex
);
298 #define define_one_rw(_name) \
299 static struct freq_attr _name = \
300 __ATTR(_name, 0644, show_##_name, store_##_name)
302 define_one_rw(sampling_rate
);
303 define_one_rw(up_threshold
);
304 define_one_rw(ignore_nice_load
);
305 define_one_rw(powersave_bias
);
307 static struct attribute
* dbs_attributes
[] = {
308 &sampling_rate_max
.attr
,
309 &sampling_rate_min
.attr
,
312 &ignore_nice_load
.attr
,
313 &powersave_bias
.attr
,
317 static struct attribute_group dbs_attr_group
= {
318 .attrs
= dbs_attributes
,
322 /************************** sysfs end ************************/
324 static void dbs_check_cpu(struct cpu_dbs_info_s
*this_dbs_info
)
326 unsigned int idle_ticks
, total_ticks
;
328 cputime64_t cur_jiffies
;
330 struct cpufreq_policy
*policy
;
333 if (!this_dbs_info
->enable
)
336 this_dbs_info
->freq_lo
= 0;
337 policy
= this_dbs_info
->cur_policy
;
338 cur_jiffies
= jiffies64_to_cputime64(get_jiffies_64());
339 total_ticks
= (unsigned int) cputime64_sub(cur_jiffies
,
340 this_dbs_info
->prev_cpu_wall
);
341 this_dbs_info
->prev_cpu_wall
= cur_jiffies
;
345 * Every sampling_rate, we check, if current idle time is less
346 * than 20% (default), then we try to increase frequency
347 * Every sampling_rate, we look for a the lowest
348 * frequency which can sustain the load while keeping idle time over
349 * 30%. If such a frequency exist, we try to decrease to this frequency.
351 * Any frequency increase takes it to the maximum frequency.
352 * Frequency reduction happens at minimum steps of
353 * 5% (default) of current frequency
357 idle_ticks
= UINT_MAX
;
358 for_each_cpu_mask(j
, policy
->cpus
) {
359 cputime64_t total_idle_ticks
;
360 unsigned int tmp_idle_ticks
;
361 struct cpu_dbs_info_s
*j_dbs_info
;
363 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
364 total_idle_ticks
= get_cpu_idle_time(j
);
365 tmp_idle_ticks
= (unsigned int) cputime64_sub(total_idle_ticks
,
366 j_dbs_info
->prev_cpu_idle
);
367 j_dbs_info
->prev_cpu_idle
= total_idle_ticks
;
369 if (tmp_idle_ticks
< idle_ticks
)
370 idle_ticks
= tmp_idle_ticks
;
372 load
= (100 * (total_ticks
- idle_ticks
)) / total_ticks
;
374 /* Check for frequency increase */
375 if (load
> dbs_tuners_ins
.up_threshold
) {
376 /* if we are already at full speed then break out early */
377 if (!dbs_tuners_ins
.powersave_bias
) {
378 if (policy
->cur
== policy
->max
)
381 __cpufreq_driver_target(policy
, policy
->max
,
384 int freq
= powersave_bias_target(policy
, policy
->max
,
386 __cpufreq_driver_target(policy
, freq
,
392 /* Check for frequency decrease */
393 /* if we cannot reduce the frequency anymore, break out early */
394 if (policy
->cur
== policy
->min
)
398 * The optimal frequency is the frequency that is the lowest that
399 * can support the current CPU usage without triggering the up
400 * policy. To be safe, we focus 10 points under the threshold.
402 if (load
< (dbs_tuners_ins
.up_threshold
- 10)) {
403 unsigned int freq_next
, freq_cur
;
405 freq_cur
= cpufreq_driver_getavg(policy
);
407 freq_cur
= policy
->cur
;
409 freq_next
= (freq_cur
* load
) /
410 (dbs_tuners_ins
.up_threshold
- 10);
412 if (!dbs_tuners_ins
.powersave_bias
) {
413 __cpufreq_driver_target(policy
, freq_next
,
416 int freq
= powersave_bias_target(policy
, freq_next
,
418 __cpufreq_driver_target(policy
, freq
,
424 static void do_dbs_timer(struct work_struct
*work
)
426 unsigned int cpu
= smp_processor_id();
427 struct cpu_dbs_info_s
*dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
428 enum dbs_sample sample_type
= dbs_info
->sample_type
;
429 /* We want all CPUs to do sampling nearly on same jiffy */
430 int delay
= usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
432 /* Permit rescheduling of this work item */
435 delay
-= jiffies
% delay
;
437 if (!dbs_info
->enable
)
439 /* Common NORMAL_SAMPLE setup */
440 dbs_info
->sample_type
= DBS_NORMAL_SAMPLE
;
441 if (!dbs_tuners_ins
.powersave_bias
||
442 sample_type
== DBS_NORMAL_SAMPLE
) {
444 dbs_check_cpu(dbs_info
);
445 unlock_cpu_hotplug();
446 if (dbs_info
->freq_lo
) {
447 /* Setup timer for SUB_SAMPLE */
448 dbs_info
->sample_type
= DBS_SUB_SAMPLE
;
449 delay
= dbs_info
->freq_hi_jiffies
;
452 __cpufreq_driver_target(dbs_info
->cur_policy
,
456 queue_delayed_work_on(cpu
, kondemand_wq
, &dbs_info
->work
, delay
);
459 static inline void dbs_timer_init(unsigned int cpu
)
461 struct cpu_dbs_info_s
*dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
462 /* We want all CPUs to do sampling nearly on same jiffy */
463 int delay
= usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
464 delay
-= jiffies
% delay
;
466 ondemand_powersave_bias_init();
467 INIT_DELAYED_WORK_NAR(&dbs_info
->work
, do_dbs_timer
);
468 dbs_info
->sample_type
= DBS_NORMAL_SAMPLE
;
469 queue_delayed_work_on(cpu
, kondemand_wq
, &dbs_info
->work
, delay
);
472 static inline void dbs_timer_exit(struct cpu_dbs_info_s
*dbs_info
)
474 dbs_info
->enable
= 0;
475 cancel_delayed_work(&dbs_info
->work
);
476 flush_workqueue(kondemand_wq
);
479 static int cpufreq_governor_dbs(struct cpufreq_policy
*policy
,
482 unsigned int cpu
= policy
->cpu
;
483 struct cpu_dbs_info_s
*this_dbs_info
;
487 this_dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
490 case CPUFREQ_GOV_START
:
491 if ((!cpu_online(cpu
)) || (!policy
->cur
))
494 if (policy
->cpuinfo
.transition_latency
>
495 (TRANSITION_LATENCY_LIMIT
* 1000)) {
496 printk(KERN_WARNING
"ondemand governor failed to load "
497 "due to too long transition latency\n");
500 if (this_dbs_info
->enable
) /* Already enabled */
503 mutex_lock(&dbs_mutex
);
505 if (dbs_enable
== 1) {
506 kondemand_wq
= create_workqueue("kondemand");
509 "Creation of kondemand failed\n");
511 mutex_unlock(&dbs_mutex
);
516 rc
= sysfs_create_group(&policy
->kobj
, &dbs_attr_group
);
519 destroy_workqueue(kondemand_wq
);
521 mutex_unlock(&dbs_mutex
);
525 for_each_cpu_mask(j
, policy
->cpus
) {
526 struct cpu_dbs_info_s
*j_dbs_info
;
527 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
528 j_dbs_info
->cur_policy
= policy
;
530 j_dbs_info
->prev_cpu_idle
= get_cpu_idle_time(j
);
531 j_dbs_info
->prev_cpu_wall
= get_jiffies_64();
533 this_dbs_info
->enable
= 1;
535 * Start the timerschedule work, when this governor
536 * is used for first time
538 if (dbs_enable
== 1) {
539 unsigned int latency
;
540 /* policy latency is in nS. Convert it to uS first */
541 latency
= policy
->cpuinfo
.transition_latency
/ 1000;
545 def_sampling_rate
= latency
*
546 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER
;
548 if (def_sampling_rate
< MIN_STAT_SAMPLING_RATE
)
549 def_sampling_rate
= MIN_STAT_SAMPLING_RATE
;
551 dbs_tuners_ins
.sampling_rate
= def_sampling_rate
;
553 dbs_timer_init(policy
->cpu
);
555 mutex_unlock(&dbs_mutex
);
558 case CPUFREQ_GOV_STOP
:
559 mutex_lock(&dbs_mutex
);
560 dbs_timer_exit(this_dbs_info
);
561 sysfs_remove_group(&policy
->kobj
, &dbs_attr_group
);
564 destroy_workqueue(kondemand_wq
);
566 mutex_unlock(&dbs_mutex
);
570 case CPUFREQ_GOV_LIMITS
:
571 mutex_lock(&dbs_mutex
);
572 if (policy
->max
< this_dbs_info
->cur_policy
->cur
)
573 __cpufreq_driver_target(this_dbs_info
->cur_policy
,
576 else if (policy
->min
> this_dbs_info
->cur_policy
->cur
)
577 __cpufreq_driver_target(this_dbs_info
->cur_policy
,
580 mutex_unlock(&dbs_mutex
);
586 static struct cpufreq_governor cpufreq_gov_dbs
= {
588 .governor
= cpufreq_governor_dbs
,
589 .owner
= THIS_MODULE
,
592 static int __init
cpufreq_gov_dbs_init(void)
594 return cpufreq_register_governor(&cpufreq_gov_dbs
);
597 static void __exit
cpufreq_gov_dbs_exit(void)
599 cpufreq_unregister_governor(&cpufreq_gov_dbs
);
603 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
604 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
605 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
606 "Low Latency Frequency Transition capable processors");
607 MODULE_LICENSE("GPL");
609 module_init(cpufreq_gov_dbs_init
);
610 module_exit(cpufreq_gov_dbs_exit
);