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/smp.h>
16 #include <linux/init.h>
17 #include <linux/interrupt.h>
18 #include <linux/ctype.h>
19 #include <linux/cpufreq.h>
20 #include <linux/sysctl.h>
21 #include <linux/types.h>
23 #include <linux/sysfs.h>
24 #include <linux/sched.h>
25 #include <linux/kmod.h>
26 #include <linux/workqueue.h>
27 #include <linux/jiffies.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/percpu.h>
32 * dbs is used in this file as a shortform for demandbased switching
33 * It helps to keep variable names smaller, simpler
36 #define DEF_FREQUENCY_UP_THRESHOLD (80)
37 #define MIN_FREQUENCY_UP_THRESHOLD (0)
38 #define MAX_FREQUENCY_UP_THRESHOLD (100)
40 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
41 #define MIN_FREQUENCY_DOWN_THRESHOLD (0)
42 #define MAX_FREQUENCY_DOWN_THRESHOLD (100)
45 * The polling frequency of this governor depends on the capability of
46 * the processor. Default polling frequency is 1000 times the transition
47 * latency of the processor. The governor will work on any processor with
48 * transition latency <= 10mS, using appropriate sampling
50 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
51 * this governor will not work.
52 * All times here are in uS.
54 static unsigned int def_sampling_rate
;
55 #define MIN_SAMPLING_RATE (def_sampling_rate / 2)
56 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
57 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
58 #define DEF_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
;
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 */
73 static DECLARE_MUTEX (dbs_sem
);
74 static DECLARE_WORK (dbs_work
, do_dbs_timer
, NULL
);
77 unsigned int sampling_rate
;
78 unsigned int sampling_down_factor
;
79 unsigned int up_threshold
;
80 unsigned int down_threshold
;
83 static struct dbs_tuners dbs_tuners_ins
= {
84 .up_threshold
= DEF_FREQUENCY_UP_THRESHOLD
,
85 .down_threshold
= DEF_FREQUENCY_DOWN_THRESHOLD
,
86 .sampling_down_factor
= DEF_SAMPLING_DOWN_FACTOR
,
89 /************************** sysfs interface ************************/
90 static ssize_t
show_sampling_rate_max(struct cpufreq_policy
*policy
, char *buf
)
92 return sprintf (buf
, "%u\n", MAX_SAMPLING_RATE
);
95 static ssize_t
show_sampling_rate_min(struct cpufreq_policy
*policy
, char *buf
)
97 return sprintf (buf
, "%u\n", MIN_SAMPLING_RATE
);
100 #define define_one_ro(_name) \
101 static struct freq_attr _name = \
102 __ATTR(_name, 0444, show_##_name, NULL)
104 define_one_ro(sampling_rate_max
);
105 define_one_ro(sampling_rate_min
);
107 /* cpufreq_ondemand Governor Tunables */
108 #define show_one(file_name, object) \
109 static ssize_t show_##file_name \
110 (struct cpufreq_policy *unused, char *buf) \
112 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
114 show_one(sampling_rate
, sampling_rate
);
115 show_one(sampling_down_factor
, sampling_down_factor
);
116 show_one(up_threshold
, up_threshold
);
117 show_one(down_threshold
, down_threshold
);
119 static ssize_t
store_sampling_down_factor(struct cpufreq_policy
*unused
,
120 const char *buf
, size_t count
)
124 ret
= sscanf (buf
, "%u", &input
);
129 dbs_tuners_ins
.sampling_down_factor
= input
;
135 static ssize_t
store_sampling_rate(struct cpufreq_policy
*unused
,
136 const char *buf
, size_t count
)
140 ret
= sscanf (buf
, "%u", &input
);
143 if (ret
!= 1 || input
> MAX_SAMPLING_RATE
|| input
< MIN_SAMPLING_RATE
) {
148 dbs_tuners_ins
.sampling_rate
= input
;
154 static ssize_t
store_up_threshold(struct cpufreq_policy
*unused
,
155 const char *buf
, size_t count
)
159 ret
= sscanf (buf
, "%u", &input
);
162 if (ret
!= 1 || input
> MAX_FREQUENCY_UP_THRESHOLD
||
163 input
< MIN_FREQUENCY_UP_THRESHOLD
||
164 input
<= dbs_tuners_ins
.down_threshold
) {
169 dbs_tuners_ins
.up_threshold
= input
;
175 static ssize_t
store_down_threshold(struct cpufreq_policy
*unused
,
176 const char *buf
, size_t count
)
180 ret
= sscanf (buf
, "%u", &input
);
183 if (ret
!= 1 || input
> MAX_FREQUENCY_DOWN_THRESHOLD
||
184 input
< MIN_FREQUENCY_DOWN_THRESHOLD
||
185 input
>= dbs_tuners_ins
.up_threshold
) {
190 dbs_tuners_ins
.down_threshold
= input
;
196 #define define_one_rw(_name) \
197 static struct freq_attr _name = \
198 __ATTR(_name, 0644, show_##_name, store_##_name)
200 define_one_rw(sampling_rate
);
201 define_one_rw(sampling_down_factor
);
202 define_one_rw(up_threshold
);
203 define_one_rw(down_threshold
);
205 static struct attribute
* dbs_attributes
[] = {
206 &sampling_rate_max
.attr
,
207 &sampling_rate_min
.attr
,
209 &sampling_down_factor
.attr
,
211 &down_threshold
.attr
,
215 static struct attribute_group dbs_attr_group
= {
216 .attrs
= dbs_attributes
,
220 /************************** sysfs end ************************/
222 static void dbs_check_cpu(int cpu
)
224 unsigned int idle_ticks
, up_idle_ticks
, down_idle_ticks
;
225 unsigned int total_idle_ticks
;
226 unsigned int freq_down_step
;
227 unsigned int freq_down_sampling_rate
;
228 static int down_skip
[NR_CPUS
];
229 struct cpu_dbs_info_s
*this_dbs_info
;
231 struct cpufreq_policy
*policy
;
234 this_dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
235 if (!this_dbs_info
->enable
)
238 policy
= this_dbs_info
->cur_policy
;
240 * The default safe range is 20% to 80%
241 * Every sampling_rate, we check
242 * - If current idle time is less than 20%, then we try to
244 * Every sampling_rate*sampling_down_factor, we check
245 * - If current idle time is more than 80%, then we try to
248 * Any frequency increase takes it to the maximum frequency.
249 * Frequency reduction happens at minimum steps of
250 * 5% of max_frequency
253 /* Check for frequency increase */
254 total_idle_ticks
= kstat_cpu(cpu
).cpustat
.idle
+
255 kstat_cpu(cpu
).cpustat
.iowait
;
256 idle_ticks
= total_idle_ticks
-
257 this_dbs_info
->prev_cpu_idle_up
;
258 this_dbs_info
->prev_cpu_idle_up
= total_idle_ticks
;
261 for_each_cpu_mask(j
, policy
->cpus
) {
262 unsigned int tmp_idle_ticks
;
263 struct cpu_dbs_info_s
*j_dbs_info
;
268 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
269 /* Check for frequency increase */
270 total_idle_ticks
= kstat_cpu(j
).cpustat
.idle
+
271 kstat_cpu(j
).cpustat
.iowait
;
272 tmp_idle_ticks
= total_idle_ticks
-
273 j_dbs_info
->prev_cpu_idle_up
;
274 j_dbs_info
->prev_cpu_idle_up
= total_idle_ticks
;
276 if (tmp_idle_ticks
< idle_ticks
)
277 idle_ticks
= tmp_idle_ticks
;
280 /* Scale idle ticks by 100 and compare with up and down ticks */
282 up_idle_ticks
= (100 - dbs_tuners_ins
.up_threshold
) *
283 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
);
285 if (idle_ticks
< up_idle_ticks
) {
286 __cpufreq_driver_target(policy
, policy
->max
,
289 this_dbs_info
->prev_cpu_idle_down
= total_idle_ticks
;
293 /* Check for frequency decrease */
295 if (down_skip
[cpu
] < dbs_tuners_ins
.sampling_down_factor
)
298 total_idle_ticks
= kstat_cpu(cpu
).cpustat
.idle
+
299 kstat_cpu(cpu
).cpustat
.iowait
;
300 idle_ticks
= total_idle_ticks
-
301 this_dbs_info
->prev_cpu_idle_down
;
302 this_dbs_info
->prev_cpu_idle_down
= total_idle_ticks
;
304 for_each_cpu_mask(j
, policy
->cpus
) {
305 unsigned int tmp_idle_ticks
;
306 struct cpu_dbs_info_s
*j_dbs_info
;
311 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
312 /* Check for frequency increase */
313 total_idle_ticks
= kstat_cpu(j
).cpustat
.idle
+
314 kstat_cpu(j
).cpustat
.iowait
;
315 tmp_idle_ticks
= total_idle_ticks
-
316 j_dbs_info
->prev_cpu_idle_down
;
317 j_dbs_info
->prev_cpu_idle_down
= total_idle_ticks
;
319 if (tmp_idle_ticks
< idle_ticks
)
320 idle_ticks
= tmp_idle_ticks
;
323 /* Scale idle ticks by 100 and compare with up and down ticks */
327 freq_down_sampling_rate
= dbs_tuners_ins
.sampling_rate
*
328 dbs_tuners_ins
.sampling_down_factor
;
329 down_idle_ticks
= (100 - dbs_tuners_ins
.down_threshold
) *
330 usecs_to_jiffies(freq_down_sampling_rate
);
332 if (idle_ticks
> down_idle_ticks
) {
333 freq_down_step
= (5 * policy
->max
) / 100;
335 /* max freq cannot be less than 100. But who knows.... */
336 if (unlikely(freq_down_step
== 0))
339 __cpufreq_driver_target(policy
,
340 policy
->cur
- freq_down_step
,
346 static void do_dbs_timer(void *data
)
350 for_each_online_cpu(i
)
352 schedule_delayed_work(&dbs_work
,
353 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
));
357 static inline void dbs_timer_init(void)
359 INIT_WORK(&dbs_work
, do_dbs_timer
, NULL
);
360 schedule_delayed_work(&dbs_work
,
361 usecs_to_jiffies(dbs_tuners_ins
.sampling_rate
));
365 static inline void dbs_timer_exit(void)
367 cancel_delayed_work(&dbs_work
);
371 static int cpufreq_governor_dbs(struct cpufreq_policy
*policy
,
374 unsigned int cpu
= policy
->cpu
;
375 struct cpu_dbs_info_s
*this_dbs_info
;
378 this_dbs_info
= &per_cpu(cpu_dbs_info
, cpu
);
381 case CPUFREQ_GOV_START
:
382 if ((!cpu_online(cpu
)) ||
386 if (policy
->cpuinfo
.transition_latency
>
387 (TRANSITION_LATENCY_LIMIT
* 1000))
389 if (this_dbs_info
->enable
) /* Already enabled */
393 for_each_cpu_mask(j
, policy
->cpus
) {
394 struct cpu_dbs_info_s
*j_dbs_info
;
395 j_dbs_info
= &per_cpu(cpu_dbs_info
, j
);
396 j_dbs_info
->cur_policy
= policy
;
398 j_dbs_info
->prev_cpu_idle_up
=
399 kstat_cpu(j
).cpustat
.idle
+
400 kstat_cpu(j
).cpustat
.iowait
;
401 j_dbs_info
->prev_cpu_idle_down
=
402 kstat_cpu(j
).cpustat
.idle
+
403 kstat_cpu(j
).cpustat
.iowait
;
405 this_dbs_info
->enable
= 1;
406 sysfs_create_group(&policy
->kobj
, &dbs_attr_group
);
409 * Start the timerschedule work, when this governor
410 * is used for first time
412 if (dbs_enable
== 1) {
413 unsigned int latency
;
414 /* policy latency is in nS. Convert it to uS first */
416 latency
= policy
->cpuinfo
.transition_latency
;
420 def_sampling_rate
= (latency
/ 1000) *
421 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER
;
422 dbs_tuners_ins
.sampling_rate
= def_sampling_rate
;
430 case CPUFREQ_GOV_STOP
:
432 this_dbs_info
->enable
= 0;
433 sysfs_remove_group(&policy
->kobj
, &dbs_attr_group
);
436 * Stop the timerschedule work, when this governor
437 * is used for first time
446 case CPUFREQ_GOV_LIMITS
:
448 if (policy
->max
< this_dbs_info
->cur_policy
->cur
)
449 __cpufreq_driver_target(
450 this_dbs_info
->cur_policy
,
451 policy
->max
, CPUFREQ_RELATION_H
);
452 else if (policy
->min
> this_dbs_info
->cur_policy
->cur
)
453 __cpufreq_driver_target(
454 this_dbs_info
->cur_policy
,
455 policy
->min
, CPUFREQ_RELATION_L
);
462 static struct cpufreq_governor cpufreq_gov_dbs
= {
464 .governor
= cpufreq_governor_dbs
,
465 .owner
= THIS_MODULE
,
468 static int __init
cpufreq_gov_dbs_init(void)
470 return cpufreq_register_governor(&cpufreq_gov_dbs
);
473 static void __exit
cpufreq_gov_dbs_exit(void)
475 /* Make sure that the scheduled work is indeed not running */
476 flush_scheduled_work();
478 cpufreq_unregister_governor(&cpufreq_gov_dbs
);
482 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
483 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
484 "Low Latency Frequency Transition capable processors");
485 MODULE_LICENSE ("GPL");
487 module_init(cpufreq_gov_dbs_init
);
488 module_exit(cpufreq_gov_dbs_exit
);