Merge master.kernel.org:/pub/scm/linux/kernel/git/davem/sparc-2.6
[pandora-kernel.git] / drivers / cpufreq / cpufreq_conservative.c
1 /*
2  *  drivers/cpufreq/cpufreq_conservative.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  *            (C)  2004 Alexander Clouter <alex-kernel@digriz.org.uk>
8  *
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.
12  */
13
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>
23 #include <linux/fs.h>
24 #include <linux/sysfs.h>
25 #include <linux/cpu.h>
26 #include <linux/sched.h>
27 #include <linux/kmod.h>
28 #include <linux/workqueue.h>
29 #include <linux/jiffies.h>
30 #include <linux/kernel_stat.h>
31 #include <linux/percpu.h>
32 #include <linux/mutex.h>
33 /*
34  * dbs is used in this file as a shortform for demandbased switching
35  * It helps to keep variable names smaller, simpler
36  */
37
38 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
39 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
40
41 /* 
42  * The polling frequency of this governor depends on the capability of 
43  * the processor. Default polling frequency is 1000 times the transition
44  * latency of the processor. The governor will work on any processor with 
45  * transition latency <= 10mS, using appropriate sampling 
46  * rate.
47  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
48  * this governor will not work.
49  * All times here are in uS.
50  */
51 static unsigned int                             def_sampling_rate;
52 #define MIN_SAMPLING_RATE_RATIO                 (2)
53 /* for correct statistics, we need at least 10 ticks between each measure */
54 #define MIN_STAT_SAMPLING_RATE                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
55 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
56 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
57 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
58 #define DEF_SAMPLING_DOWN_FACTOR                (1)
59 #define MAX_SAMPLING_DOWN_FACTOR                (10)
60 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
61
62 static void do_dbs_timer(void *data);
63
64 struct cpu_dbs_info_s {
65         struct cpufreq_policy   *cur_policy;
66         unsigned int            prev_cpu_idle_up;
67         unsigned int            prev_cpu_idle_down;
68         unsigned int            enable;
69         unsigned int            down_skip;
70         unsigned int            requested_freq;
71 };
72 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
73
74 static unsigned int dbs_enable; /* number of CPUs using this policy */
75
76 /*
77  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
78  * lock and dbs_mutex. cpu_hotplug lock should always be held before
79  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
80  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
81  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
82  * is recursive for the same process. -Venki
83  */
84 static DEFINE_MUTEX     (dbs_mutex);
85 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
86
87 struct dbs_tuners {
88         unsigned int            sampling_rate;
89         unsigned int            sampling_down_factor;
90         unsigned int            up_threshold;
91         unsigned int            down_threshold;
92         unsigned int            ignore_nice;
93         unsigned int            freq_step;
94 };
95
96 static struct dbs_tuners dbs_tuners_ins = {
97         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
98         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
99         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
100         .ignore_nice            = 0,
101         .freq_step              = 5,
102 };
103
104 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
105 {
106         return  kstat_cpu(cpu).cpustat.idle +
107                 kstat_cpu(cpu).cpustat.iowait +
108                 ( dbs_tuners_ins.ignore_nice ?
109                   kstat_cpu(cpu).cpustat.nice :
110                   0);
111 }
112
113 /************************** sysfs interface ************************/
114 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
115 {
116         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
117 }
118
119 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
120 {
121         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
122 }
123
124 #define define_one_ro(_name)                                    \
125 static struct freq_attr _name =                                 \
126 __ATTR(_name, 0444, show_##_name, NULL)
127
128 define_one_ro(sampling_rate_max);
129 define_one_ro(sampling_rate_min);
130
131 /* cpufreq_conservative Governor Tunables */
132 #define show_one(file_name, object)                                     \
133 static ssize_t show_##file_name                                         \
134 (struct cpufreq_policy *unused, char *buf)                              \
135 {                                                                       \
136         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
137 }
138 show_one(sampling_rate, sampling_rate);
139 show_one(sampling_down_factor, sampling_down_factor);
140 show_one(up_threshold, up_threshold);
141 show_one(down_threshold, down_threshold);
142 show_one(ignore_nice_load, ignore_nice);
143 show_one(freq_step, freq_step);
144
145 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
146                 const char *buf, size_t count)
147 {
148         unsigned int input;
149         int ret;
150         ret = sscanf (buf, "%u", &input);
151         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
152                 return -EINVAL;
153
154         mutex_lock(&dbs_mutex);
155         dbs_tuners_ins.sampling_down_factor = input;
156         mutex_unlock(&dbs_mutex);
157
158         return count;
159 }
160
161 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
162                 const char *buf, size_t count)
163 {
164         unsigned int input;
165         int ret;
166         ret = sscanf (buf, "%u", &input);
167
168         mutex_lock(&dbs_mutex);
169         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
170                 mutex_unlock(&dbs_mutex);
171                 return -EINVAL;
172         }
173
174         dbs_tuners_ins.sampling_rate = input;
175         mutex_unlock(&dbs_mutex);
176
177         return count;
178 }
179
180 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
181                 const char *buf, size_t count)
182 {
183         unsigned int input;
184         int ret;
185         ret = sscanf (buf, "%u", &input);
186
187         mutex_lock(&dbs_mutex);
188         if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) {
189                 mutex_unlock(&dbs_mutex);
190                 return -EINVAL;
191         }
192
193         dbs_tuners_ins.up_threshold = input;
194         mutex_unlock(&dbs_mutex);
195
196         return count;
197 }
198
199 static ssize_t store_down_threshold(struct cpufreq_policy *unused, 
200                 const char *buf, size_t count)
201 {
202         unsigned int input;
203         int ret;
204         ret = sscanf (buf, "%u", &input);
205
206         mutex_lock(&dbs_mutex);
207         if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
208                 mutex_unlock(&dbs_mutex);
209                 return -EINVAL;
210         }
211
212         dbs_tuners_ins.down_threshold = input;
213         mutex_unlock(&dbs_mutex);
214
215         return count;
216 }
217
218 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
219                 const char *buf, size_t count)
220 {
221         unsigned int input;
222         int ret;
223
224         unsigned int j;
225         
226         ret = sscanf (buf, "%u", &input);
227         if ( ret != 1 )
228                 return -EINVAL;
229
230         if ( input > 1 )
231                 input = 1;
232         
233         mutex_lock(&dbs_mutex);
234         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
235                 mutex_unlock(&dbs_mutex);
236                 return count;
237         }
238         dbs_tuners_ins.ignore_nice = input;
239
240         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
241         for_each_online_cpu(j) {
242                 struct cpu_dbs_info_s *j_dbs_info;
243                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
244                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
245                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
246         }
247         mutex_unlock(&dbs_mutex);
248
249         return count;
250 }
251
252 static ssize_t store_freq_step(struct cpufreq_policy *policy,
253                 const char *buf, size_t count)
254 {
255         unsigned int input;
256         int ret;
257
258         ret = sscanf (buf, "%u", &input);
259
260         if ( ret != 1 )
261                 return -EINVAL;
262
263         if ( input > 100 )
264                 input = 100;
265         
266         /* no need to test here if freq_step is zero as the user might actually
267          * want this, they would be crazy though :) */
268         mutex_lock(&dbs_mutex);
269         dbs_tuners_ins.freq_step = input;
270         mutex_unlock(&dbs_mutex);
271
272         return count;
273 }
274
275 #define define_one_rw(_name) \
276 static struct freq_attr _name = \
277 __ATTR(_name, 0644, show_##_name, store_##_name)
278
279 define_one_rw(sampling_rate);
280 define_one_rw(sampling_down_factor);
281 define_one_rw(up_threshold);
282 define_one_rw(down_threshold);
283 define_one_rw(ignore_nice_load);
284 define_one_rw(freq_step);
285
286 static struct attribute * dbs_attributes[] = {
287         &sampling_rate_max.attr,
288         &sampling_rate_min.attr,
289         &sampling_rate.attr,
290         &sampling_down_factor.attr,
291         &up_threshold.attr,
292         &down_threshold.attr,
293         &ignore_nice_load.attr,
294         &freq_step.attr,
295         NULL
296 };
297
298 static struct attribute_group dbs_attr_group = {
299         .attrs = dbs_attributes,
300         .name = "conservative",
301 };
302
303 /************************** sysfs end ************************/
304
305 static void dbs_check_cpu(int cpu)
306 {
307         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
308         unsigned int tmp_idle_ticks, total_idle_ticks;
309         unsigned int freq_step;
310         unsigned int freq_down_sampling_rate;
311         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
312         struct cpufreq_policy *policy;
313
314         if (!this_dbs_info->enable)
315                 return;
316
317         policy = this_dbs_info->cur_policy;
318
319         /* 
320          * The default safe range is 20% to 80% 
321          * Every sampling_rate, we check
322          *      - If current idle time is less than 20%, then we try to 
323          *        increase frequency
324          * Every sampling_rate*sampling_down_factor, we check
325          *      - If current idle time is more than 80%, then we try to
326          *        decrease frequency
327          *
328          * Any frequency increase takes it to the maximum frequency. 
329          * Frequency reduction happens at minimum steps of 
330          * 5% (default) of max_frequency 
331          */
332
333         /* Check for frequency increase */
334         idle_ticks = UINT_MAX;
335
336         /* Check for frequency increase */
337         total_idle_ticks = get_cpu_idle_time(cpu);
338         tmp_idle_ticks = total_idle_ticks -
339                 this_dbs_info->prev_cpu_idle_up;
340         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
341
342         if (tmp_idle_ticks < idle_ticks)
343                 idle_ticks = tmp_idle_ticks;
344
345         /* Scale idle ticks by 100 and compare with up and down ticks */
346         idle_ticks *= 100;
347         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
348                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
349
350         if (idle_ticks < up_idle_ticks) {
351                 this_dbs_info->down_skip = 0;
352                 this_dbs_info->prev_cpu_idle_down =
353                         this_dbs_info->prev_cpu_idle_up;
354
355                 /* if we are already at full speed then break out early */
356                 if (this_dbs_info->requested_freq == policy->max)
357                         return;
358                 
359                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
360
361                 /* max freq cannot be less than 100. But who knows.... */
362                 if (unlikely(freq_step == 0))
363                         freq_step = 5;
364                 
365                 this_dbs_info->requested_freq += freq_step;
366                 if (this_dbs_info->requested_freq > policy->max)
367                         this_dbs_info->requested_freq = policy->max;
368
369                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
370                         CPUFREQ_RELATION_H);
371                 return;
372         }
373
374         /* Check for frequency decrease */
375         this_dbs_info->down_skip++;
376         if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
377                 return;
378
379         /* Check for frequency decrease */
380         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
381         tmp_idle_ticks = total_idle_ticks -
382                 this_dbs_info->prev_cpu_idle_down;
383         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
384
385         if (tmp_idle_ticks < idle_ticks)
386                 idle_ticks = tmp_idle_ticks;
387
388         /* Scale idle ticks by 100 and compare with up and down ticks */
389         idle_ticks *= 100;
390         this_dbs_info->down_skip = 0;
391
392         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
393                 dbs_tuners_ins.sampling_down_factor;
394         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
395                 usecs_to_jiffies(freq_down_sampling_rate);
396
397         if (idle_ticks > down_idle_ticks) {
398                 /*
399                  * if we are already at the lowest speed then break out early
400                  * or if we 'cannot' reduce the speed as the user might want
401                  * freq_step to be zero
402                  */
403                 if (this_dbs_info->requested_freq == policy->min
404                                 || dbs_tuners_ins.freq_step == 0)
405                         return;
406
407                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
408
409                 /* max freq cannot be less than 100. But who knows.... */
410                 if (unlikely(freq_step == 0))
411                         freq_step = 5;
412
413                 this_dbs_info->requested_freq -= freq_step;
414                 if (this_dbs_info->requested_freq < policy->min)
415                         this_dbs_info->requested_freq = policy->min;
416
417                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
418                                 CPUFREQ_RELATION_H);
419                 return;
420         }
421 }
422
423 static void do_dbs_timer(void *data)
424
425         int i;
426         lock_cpu_hotplug();
427         mutex_lock(&dbs_mutex);
428         for_each_online_cpu(i)
429                 dbs_check_cpu(i);
430         schedule_delayed_work(&dbs_work, 
431                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
432         mutex_unlock(&dbs_mutex);
433         unlock_cpu_hotplug();
434
435
436 static inline void dbs_timer_init(void)
437 {
438         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
439         schedule_delayed_work(&dbs_work,
440                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
441         return;
442 }
443
444 static inline void dbs_timer_exit(void)
445 {
446         cancel_delayed_work(&dbs_work);
447         return;
448 }
449
450 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
451                                    unsigned int event)
452 {
453         unsigned int cpu = policy->cpu;
454         struct cpu_dbs_info_s *this_dbs_info;
455         unsigned int j;
456
457         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
458
459         switch (event) {
460         case CPUFREQ_GOV_START:
461                 if ((!cpu_online(cpu)) || 
462                     (!policy->cur))
463                         return -EINVAL;
464
465                 if (policy->cpuinfo.transition_latency >
466                                 (TRANSITION_LATENCY_LIMIT * 1000))
467                         return -EINVAL;
468                 if (this_dbs_info->enable) /* Already enabled */
469                         break;
470                  
471                 mutex_lock(&dbs_mutex);
472                 for_each_cpu_mask(j, policy->cpus) {
473                         struct cpu_dbs_info_s *j_dbs_info;
474                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
475                         j_dbs_info->cur_policy = policy;
476                 
477                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
478                         j_dbs_info->prev_cpu_idle_down
479                                 = j_dbs_info->prev_cpu_idle_up;
480                 }
481                 this_dbs_info->enable = 1;
482                 this_dbs_info->down_skip = 0;
483                 this_dbs_info->requested_freq = policy->cur;
484                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
485                 dbs_enable++;
486                 /*
487                  * Start the timerschedule work, when this governor
488                  * is used for first time
489                  */
490                 if (dbs_enable == 1) {
491                         unsigned int latency;
492                         /* policy latency is in nS. Convert it to uS first */
493                         latency = policy->cpuinfo.transition_latency / 1000;
494                         if (latency == 0)
495                                 latency = 1;
496
497                         def_sampling_rate = 10 * latency *
498                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
499
500                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
501                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
502
503                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
504
505                         dbs_timer_init();
506                 }
507                 
508                 mutex_unlock(&dbs_mutex);
509                 break;
510
511         case CPUFREQ_GOV_STOP:
512                 mutex_lock(&dbs_mutex);
513                 this_dbs_info->enable = 0;
514                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
515                 dbs_enable--;
516                 /*
517                  * Stop the timerschedule work, when this governor
518                  * is used for first time
519                  */
520                 if (dbs_enable == 0) 
521                         dbs_timer_exit();
522                 
523                 mutex_unlock(&dbs_mutex);
524
525                 break;
526
527         case CPUFREQ_GOV_LIMITS:
528                 mutex_lock(&dbs_mutex);
529                 if (policy->max < this_dbs_info->cur_policy->cur)
530                         __cpufreq_driver_target(
531                                         this_dbs_info->cur_policy,
532                                         policy->max, CPUFREQ_RELATION_H);
533                 else if (policy->min > this_dbs_info->cur_policy->cur)
534                         __cpufreq_driver_target(
535                                         this_dbs_info->cur_policy,
536                                         policy->min, CPUFREQ_RELATION_L);
537                 mutex_unlock(&dbs_mutex);
538                 break;
539         }
540         return 0;
541 }
542
543 static struct cpufreq_governor cpufreq_gov_dbs = {
544         .name           = "conservative",
545         .governor       = cpufreq_governor_dbs,
546         .owner          = THIS_MODULE,
547 };
548
549 static int __init cpufreq_gov_dbs_init(void)
550 {
551         return cpufreq_register_governor(&cpufreq_gov_dbs);
552 }
553
554 static void __exit cpufreq_gov_dbs_exit(void)
555 {
556         /* Make sure that the scheduled work is indeed not running */
557         flush_scheduled_work();
558
559         cpufreq_unregister_governor(&cpufreq_gov_dbs);
560 }
561
562
563 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
564 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
565                 "Low Latency Frequency Transition capable processors "
566                 "optimised for use in a battery environment");
567 MODULE_LICENSE ("GPL");
568
569 module_init(cpufreq_gov_dbs_init);
570 module_exit(cpufreq_gov_dbs_exit);