Merge master.kernel.org:/home/rmk/linux-2.6-arm
[linux-2.6] / drivers / cpufreq / cpufreq_ondemand.c
1 /*
2  *  drivers/cpufreq/cpufreq_ondemand.c
3  *
4  *  Copyright (C)  2001 Russell King
5  *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6  *                      Jun Nakajima <jun.nakajima@intel.com>
7  *
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.
11  */
12
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>
22 #include <linux/fs.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>
30
31 /*
32  * dbs is used in this file as a shortform for demandbased switching
33  * It helps to keep variable names smaller, simpler
34  */
35
36 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
37 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
38 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
39
40 /* 
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 
45  * rate.
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.
49  */
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)
60
61 static void do_dbs_timer(void *data);
62
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;
67         unsigned int            enable;
68 };
69 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
70
71 static unsigned int dbs_enable; /* number of CPUs using this policy */
72
73 static DECLARE_MUTEX    (dbs_sem);
74 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
75
76 struct dbs_tuners {
77         unsigned int            sampling_rate;
78         unsigned int            sampling_down_factor;
79         unsigned int            up_threshold;
80         unsigned int            ignore_nice;
81 };
82
83 static struct dbs_tuners dbs_tuners_ins = {
84         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
85         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
86 };
87
88 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
89 {
90         return  kstat_cpu(cpu).cpustat.idle +
91                 kstat_cpu(cpu).cpustat.iowait +
92                 ( !dbs_tuners_ins.ignore_nice ? 
93                   kstat_cpu(cpu).cpustat.nice :
94                   0);
95 }
96
97 /************************** sysfs interface ************************/
98 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
99 {
100         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
101 }
102
103 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
104 {
105         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
106 }
107
108 #define define_one_ro(_name)                                    \
109 static struct freq_attr _name =                                 \
110 __ATTR(_name, 0444, show_##_name, NULL)
111
112 define_one_ro(sampling_rate_max);
113 define_one_ro(sampling_rate_min);
114
115 /* cpufreq_ondemand Governor Tunables */
116 #define show_one(file_name, object)                                     \
117 static ssize_t show_##file_name                                         \
118 (struct cpufreq_policy *unused, char *buf)                              \
119 {                                                                       \
120         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
121 }
122 show_one(sampling_rate, sampling_rate);
123 show_one(sampling_down_factor, sampling_down_factor);
124 show_one(up_threshold, up_threshold);
125 show_one(ignore_nice, ignore_nice);
126
127 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
128                 const char *buf, size_t count)
129 {
130         unsigned int input;
131         int ret;
132         ret = sscanf (buf, "%u", &input);
133         if (ret != 1 )
134                 return -EINVAL;
135
136         if (input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
137                 return -EINVAL;
138
139         down(&dbs_sem);
140         dbs_tuners_ins.sampling_down_factor = input;
141         up(&dbs_sem);
142
143         return count;
144 }
145
146 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
147                 const char *buf, size_t count)
148 {
149         unsigned int input;
150         int ret;
151         ret = sscanf (buf, "%u", &input);
152
153         down(&dbs_sem);
154         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
155                 up(&dbs_sem);
156                 return -EINVAL;
157         }
158
159         dbs_tuners_ins.sampling_rate = input;
160         up(&dbs_sem);
161
162         return count;
163 }
164
165 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
166                 const char *buf, size_t count)
167 {
168         unsigned int input;
169         int ret;
170         ret = sscanf (buf, "%u", &input);
171
172         down(&dbs_sem);
173         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 
174                         input < MIN_FREQUENCY_UP_THRESHOLD) {
175                 up(&dbs_sem);
176                 return -EINVAL;
177         }
178
179         dbs_tuners_ins.up_threshold = input;
180         up(&dbs_sem);
181
182         return count;
183 }
184
185 static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
186                 const char *buf, size_t count)
187 {
188         unsigned int input;
189         int ret;
190
191         unsigned int j;
192         
193         ret = sscanf (buf, "%u", &input);
194         if ( ret != 1 )
195                 return -EINVAL;
196
197         if ( input > 1 )
198                 input = 1;
199         
200         down(&dbs_sem);
201         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
202                 up(&dbs_sem);
203                 return count;
204         }
205         dbs_tuners_ins.ignore_nice = input;
206
207         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
208         for_each_online_cpu(j) {
209                 struct cpu_dbs_info_s *j_dbs_info;
210                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
211                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
212                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
213         }
214         up(&dbs_sem);
215
216         return count;
217 }
218
219 #define define_one_rw(_name) \
220 static struct freq_attr _name = \
221 __ATTR(_name, 0644, show_##_name, store_##_name)
222
223 define_one_rw(sampling_rate);
224 define_one_rw(sampling_down_factor);
225 define_one_rw(up_threshold);
226 define_one_rw(ignore_nice);
227
228 static struct attribute * dbs_attributes[] = {
229         &sampling_rate_max.attr,
230         &sampling_rate_min.attr,
231         &sampling_rate.attr,
232         &sampling_down_factor.attr,
233         &up_threshold.attr,
234         &ignore_nice.attr,
235         NULL
236 };
237
238 static struct attribute_group dbs_attr_group = {
239         .attrs = dbs_attributes,
240         .name = "ondemand",
241 };
242
243 /************************** sysfs end ************************/
244
245 static void dbs_check_cpu(int cpu)
246 {
247         unsigned int idle_ticks, up_idle_ticks, total_ticks;
248         unsigned int freq_next;
249         unsigned int freq_down_sampling_rate;
250         static int down_skip[NR_CPUS];
251         struct cpu_dbs_info_s *this_dbs_info;
252
253         struct cpufreq_policy *policy;
254         unsigned int j;
255
256         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
257         if (!this_dbs_info->enable)
258                 return;
259
260         policy = this_dbs_info->cur_policy;
261         /* 
262          * Every sampling_rate, we check, if current idle time is less
263          * than 20% (default), then we try to increase frequency
264          * Every sampling_rate*sampling_down_factor, we look for a the lowest
265          * frequency which can sustain the load while keeping idle time over
266          * 30%. If such a frequency exist, we try to decrease to this frequency.
267          *
268          * Any frequency increase takes it to the maximum frequency. 
269          * Frequency reduction happens at minimum steps of 
270          * 5% (default) of current frequency 
271          */
272
273         /* Check for frequency increase */
274         idle_ticks = UINT_MAX;
275         for_each_cpu_mask(j, policy->cpus) {
276                 unsigned int tmp_idle_ticks, total_idle_ticks;
277                 struct cpu_dbs_info_s *j_dbs_info;
278
279                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
280                 total_idle_ticks = get_cpu_idle_time(j);
281                 tmp_idle_ticks = total_idle_ticks -
282                         j_dbs_info->prev_cpu_idle_up;
283                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
284
285                 if (tmp_idle_ticks < idle_ticks)
286                         idle_ticks = tmp_idle_ticks;
287         }
288
289         /* Scale idle ticks by 100 and compare with up and down ticks */
290         idle_ticks *= 100;
291         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
292                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
293
294         if (idle_ticks < up_idle_ticks) {
295                 down_skip[cpu] = 0;
296                 for_each_cpu_mask(j, policy->cpus) {
297                         struct cpu_dbs_info_s *j_dbs_info;
298
299                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
300                         j_dbs_info->prev_cpu_idle_down = 
301                                         j_dbs_info->prev_cpu_idle_up;
302                 }
303                 /* if we are already at full speed then break out early */
304                 if (policy->cur == policy->max)
305                         return;
306                 
307                 __cpufreq_driver_target(policy, policy->max, 
308                         CPUFREQ_RELATION_H);
309                 return;
310         }
311
312         /* Check for frequency decrease */
313         down_skip[cpu]++;
314         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
315                 return;
316
317         idle_ticks = UINT_MAX;
318         for_each_cpu_mask(j, policy->cpus) {
319                 unsigned int tmp_idle_ticks, total_idle_ticks;
320                 struct cpu_dbs_info_s *j_dbs_info;
321
322                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
323                 /* Check for frequency decrease */
324                 total_idle_ticks = j_dbs_info->prev_cpu_idle_up;
325                 tmp_idle_ticks = total_idle_ticks -
326                         j_dbs_info->prev_cpu_idle_down;
327                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
328
329                 if (tmp_idle_ticks < idle_ticks)
330                         idle_ticks = tmp_idle_ticks;
331         }
332
333         down_skip[cpu] = 0;
334         /* if we cannot reduce the frequency anymore, break out early */
335         if (policy->cur == policy->min)
336                 return;
337
338         /* Compute how many ticks there are between two measurements */
339         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
340                 dbs_tuners_ins.sampling_down_factor;
341         total_ticks = usecs_to_jiffies(freq_down_sampling_rate);
342
343         /*
344          * The optimal frequency is the frequency that is the lowest that
345          * can support the current CPU usage without triggering the up
346          * policy. To be safe, we focus 10 points under the threshold.
347          */
348         freq_next = ((total_ticks - idle_ticks) * 100) / total_ticks;
349         freq_next = (freq_next * policy->cur) / 
350                         (dbs_tuners_ins.up_threshold - 10);
351
352         if (freq_next <= ((policy->cur * 95) / 100))
353                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
354 }
355
356 static void do_dbs_timer(void *data)
357
358         int i;
359         down(&dbs_sem);
360         for_each_online_cpu(i)
361                 dbs_check_cpu(i);
362         schedule_delayed_work(&dbs_work, 
363                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
364         up(&dbs_sem);
365
366
367 static inline void dbs_timer_init(void)
368 {
369         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
370         schedule_delayed_work(&dbs_work,
371                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
372         return;
373 }
374
375 static inline void dbs_timer_exit(void)
376 {
377         cancel_delayed_work(&dbs_work);
378         return;
379 }
380
381 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
382                                    unsigned int event)
383 {
384         unsigned int cpu = policy->cpu;
385         struct cpu_dbs_info_s *this_dbs_info;
386         unsigned int j;
387
388         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
389
390         switch (event) {
391         case CPUFREQ_GOV_START:
392                 if ((!cpu_online(cpu)) || 
393                     (!policy->cur))
394                         return -EINVAL;
395
396                 if (policy->cpuinfo.transition_latency >
397                                 (TRANSITION_LATENCY_LIMIT * 1000))
398                         return -EINVAL;
399                 if (this_dbs_info->enable) /* Already enabled */
400                         break;
401                  
402                 down(&dbs_sem);
403                 for_each_cpu_mask(j, policy->cpus) {
404                         struct cpu_dbs_info_s *j_dbs_info;
405                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
406                         j_dbs_info->cur_policy = policy;
407                 
408                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
409                         j_dbs_info->prev_cpu_idle_down
410                                 = j_dbs_info->prev_cpu_idle_up;
411                 }
412                 this_dbs_info->enable = 1;
413                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
414                 dbs_enable++;
415                 /*
416                  * Start the timerschedule work, when this governor
417                  * is used for first time
418                  */
419                 if (dbs_enable == 1) {
420                         unsigned int latency;
421                         /* policy latency is in nS. Convert it to uS first */
422                         latency = policy->cpuinfo.transition_latency / 1000;
423                         if (latency == 0)
424                                 latency = 1;
425
426                         def_sampling_rate = latency *
427                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
428
429                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
430                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
431
432                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
433                         dbs_tuners_ins.ignore_nice = 0;
434
435                         dbs_timer_init();
436                 }
437                 
438                 up(&dbs_sem);
439                 break;
440
441         case CPUFREQ_GOV_STOP:
442                 down(&dbs_sem);
443                 this_dbs_info->enable = 0;
444                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
445                 dbs_enable--;
446                 /*
447                  * Stop the timerschedule work, when this governor
448                  * is used for first time
449                  */
450                 if (dbs_enable == 0) 
451                         dbs_timer_exit();
452                 
453                 up(&dbs_sem);
454
455                 break;
456
457         case CPUFREQ_GOV_LIMITS:
458                 down(&dbs_sem);
459                 if (policy->max < this_dbs_info->cur_policy->cur)
460                         __cpufreq_driver_target(
461                                         this_dbs_info->cur_policy,
462                                         policy->max, CPUFREQ_RELATION_H);
463                 else if (policy->min > this_dbs_info->cur_policy->cur)
464                         __cpufreq_driver_target(
465                                         this_dbs_info->cur_policy,
466                                         policy->min, CPUFREQ_RELATION_L);
467                 up(&dbs_sem);
468                 break;
469         }
470         return 0;
471 }
472
473 static struct cpufreq_governor cpufreq_gov_dbs = {
474         .name           = "ondemand",
475         .governor       = cpufreq_governor_dbs,
476         .owner          = THIS_MODULE,
477 };
478
479 static int __init cpufreq_gov_dbs_init(void)
480 {
481         return cpufreq_register_governor(&cpufreq_gov_dbs);
482 }
483
484 static void __exit cpufreq_gov_dbs_exit(void)
485 {
486         /* Make sure that the scheduled work is indeed not running */
487         flush_scheduled_work();
488
489         cpufreq_unregister_governor(&cpufreq_gov_dbs);
490 }
491
492
493 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
494 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
495                 "Low Latency Frequency Transition capable processors");
496 MODULE_LICENSE ("GPL");
497
498 module_init(cpufreq_gov_dbs_init);
499 module_exit(cpufreq_gov_dbs_exit);