Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/cpufreq
[linux-2.6.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/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>
32 /*
33  * dbs is used in this file as a shortform for demandbased switching
34  * It helps to keep variable names smaller, simpler
35  */
36
37 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
38 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
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
47  * with CPUFREQ_ETERNAL), 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                  \
54                         (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
55 #define MIN_SAMPLING_RATE                       \
56                         (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
57 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
58 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
59 #define DEF_SAMPLING_DOWN_FACTOR                (1)
60 #define MAX_SAMPLING_DOWN_FACTOR                (10)
61 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
62
63 static void do_dbs_timer(struct work_struct *work);
64
65 struct cpu_dbs_info_s {
66         struct cpufreq_policy   *cur_policy;
67         unsigned int            prev_cpu_idle_up;
68         unsigned int            prev_cpu_idle_down;
69         unsigned int            enable;
70         unsigned int            down_skip;
71         unsigned int            requested_freq;
72 };
73 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
74
75 static unsigned int dbs_enable; /* number of CPUs using this policy */
76
77 /*
78  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
79  * lock and dbs_mutex. cpu_hotplug lock should always be held before
80  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
81  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
82  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
83  * is recursive for the same process. -Venki
84  */
85 static DEFINE_MUTEX     (dbs_mutex);
86 static DECLARE_DELAYED_WORK(dbs_work, do_dbs_timer);
87
88 struct dbs_tuners {
89         unsigned int            sampling_rate;
90         unsigned int            sampling_down_factor;
91         unsigned int            up_threshold;
92         unsigned int            down_threshold;
93         unsigned int            ignore_nice;
94         unsigned int            freq_step;
95 };
96
97 static struct dbs_tuners dbs_tuners_ins = {
98         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
99         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
100         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
101         .ignore_nice            = 0,
102         .freq_step              = 5,
103 };
104
105 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
106 {
107         unsigned int add_nice = 0, ret;
108
109         if (dbs_tuners_ins.ignore_nice)
110                 add_nice = kstat_cpu(cpu).cpustat.nice;
111
112         ret =   kstat_cpu(cpu).cpustat.idle +
113                 kstat_cpu(cpu).cpustat.iowait +
114                 add_nice;
115
116         return ret;
117 }
118
119 /************************** sysfs interface ************************/
120 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
121 {
122         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
123 }
124
125 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
126 {
127         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
128 }
129
130 #define define_one_ro(_name)                                    \
131 static struct freq_attr _name =                                 \
132 __ATTR(_name, 0444, show_##_name, NULL)
133
134 define_one_ro(sampling_rate_max);
135 define_one_ro(sampling_rate_min);
136
137 /* cpufreq_conservative Governor Tunables */
138 #define show_one(file_name, object)                                     \
139 static ssize_t show_##file_name                                         \
140 (struct cpufreq_policy *unused, char *buf)                              \
141 {                                                                       \
142         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
143 }
144 show_one(sampling_rate, sampling_rate);
145 show_one(sampling_down_factor, sampling_down_factor);
146 show_one(up_threshold, up_threshold);
147 show_one(down_threshold, down_threshold);
148 show_one(ignore_nice_load, ignore_nice);
149 show_one(freq_step, freq_step);
150
151 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
152                 const char *buf, size_t count)
153 {
154         unsigned int input;
155         int ret;
156         ret = sscanf (buf, "%u", &input);
157         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
158                 return -EINVAL;
159
160         mutex_lock(&dbs_mutex);
161         dbs_tuners_ins.sampling_down_factor = input;
162         mutex_unlock(&dbs_mutex);
163
164         return count;
165 }
166
167 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
168                 const char *buf, size_t count)
169 {
170         unsigned int input;
171         int ret;
172         ret = sscanf (buf, "%u", &input);
173
174         mutex_lock(&dbs_mutex);
175         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
176                 mutex_unlock(&dbs_mutex);
177                 return -EINVAL;
178         }
179
180         dbs_tuners_ins.sampling_rate = input;
181         mutex_unlock(&dbs_mutex);
182
183         return count;
184 }
185
186 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
187                 const char *buf, size_t count)
188 {
189         unsigned int input;
190         int ret;
191         ret = sscanf (buf, "%u", &input);
192
193         mutex_lock(&dbs_mutex);
194         if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) {
195                 mutex_unlock(&dbs_mutex);
196                 return -EINVAL;
197         }
198
199         dbs_tuners_ins.up_threshold = input;
200         mutex_unlock(&dbs_mutex);
201
202         return count;
203 }
204
205 static ssize_t store_down_threshold(struct cpufreq_policy *unused, 
206                 const char *buf, size_t count)
207 {
208         unsigned int input;
209         int ret;
210         ret = sscanf (buf, "%u", &input);
211
212         mutex_lock(&dbs_mutex);
213         if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
214                 mutex_unlock(&dbs_mutex);
215                 return -EINVAL;
216         }
217
218         dbs_tuners_ins.down_threshold = input;
219         mutex_unlock(&dbs_mutex);
220
221         return count;
222 }
223
224 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
225                 const char *buf, size_t count)
226 {
227         unsigned int input;
228         int ret;
229
230         unsigned int j;
231         
232         ret = sscanf (buf, "%u", &input);
233         if ( ret != 1 )
234                 return -EINVAL;
235
236         if ( input > 1 )
237                 input = 1;
238         
239         mutex_lock(&dbs_mutex);
240         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
241                 mutex_unlock(&dbs_mutex);
242                 return count;
243         }
244         dbs_tuners_ins.ignore_nice = input;
245
246         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
247         for_each_online_cpu(j) {
248                 struct cpu_dbs_info_s *j_dbs_info;
249                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
250                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
251                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
252         }
253         mutex_unlock(&dbs_mutex);
254
255         return count;
256 }
257
258 static ssize_t store_freq_step(struct cpufreq_policy *policy,
259                 const char *buf, size_t count)
260 {
261         unsigned int input;
262         int ret;
263
264         ret = sscanf (buf, "%u", &input);
265
266         if ( ret != 1 )
267                 return -EINVAL;
268
269         if ( input > 100 )
270                 input = 100;
271         
272         /* no need to test here if freq_step is zero as the user might actually
273          * want this, they would be crazy though :) */
274         mutex_lock(&dbs_mutex);
275         dbs_tuners_ins.freq_step = input;
276         mutex_unlock(&dbs_mutex);
277
278         return count;
279 }
280
281 #define define_one_rw(_name) \
282 static struct freq_attr _name = \
283 __ATTR(_name, 0644, show_##_name, store_##_name)
284
285 define_one_rw(sampling_rate);
286 define_one_rw(sampling_down_factor);
287 define_one_rw(up_threshold);
288 define_one_rw(down_threshold);
289 define_one_rw(ignore_nice_load);
290 define_one_rw(freq_step);
291
292 static struct attribute * dbs_attributes[] = {
293         &sampling_rate_max.attr,
294         &sampling_rate_min.attr,
295         &sampling_rate.attr,
296         &sampling_down_factor.attr,
297         &up_threshold.attr,
298         &down_threshold.attr,
299         &ignore_nice_load.attr,
300         &freq_step.attr,
301         NULL
302 };
303
304 static struct attribute_group dbs_attr_group = {
305         .attrs = dbs_attributes,
306         .name = "conservative",
307 };
308
309 /************************** sysfs end ************************/
310
311 static void dbs_check_cpu(int cpu)
312 {
313         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
314         unsigned int tmp_idle_ticks, total_idle_ticks;
315         unsigned int freq_step;
316         unsigned int freq_down_sampling_rate;
317         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
318         struct cpufreq_policy *policy;
319
320         if (!this_dbs_info->enable)
321                 return;
322
323         policy = this_dbs_info->cur_policy;
324
325         /* 
326          * The default safe range is 20% to 80% 
327          * Every sampling_rate, we check
328          *      - If current idle time is less than 20%, then we try to 
329          *        increase frequency
330          * Every sampling_rate*sampling_down_factor, we check
331          *      - If current idle time is more than 80%, then we try to
332          *        decrease frequency
333          *
334          * Any frequency increase takes it to the maximum frequency. 
335          * Frequency reduction happens at minimum steps of 
336          * 5% (default) of max_frequency 
337          */
338
339         /* Check for frequency increase */
340         idle_ticks = UINT_MAX;
341
342         /* Check for frequency increase */
343         total_idle_ticks = get_cpu_idle_time(cpu);
344         tmp_idle_ticks = total_idle_ticks -
345                 this_dbs_info->prev_cpu_idle_up;
346         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
347
348         if (tmp_idle_ticks < idle_ticks)
349                 idle_ticks = tmp_idle_ticks;
350
351         /* Scale idle ticks by 100 and compare with up and down ticks */
352         idle_ticks *= 100;
353         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
354                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
355
356         if (idle_ticks < up_idle_ticks) {
357                 this_dbs_info->down_skip = 0;
358                 this_dbs_info->prev_cpu_idle_down =
359                         this_dbs_info->prev_cpu_idle_up;
360
361                 /* if we are already at full speed then break out early */
362                 if (this_dbs_info->requested_freq == policy->max)
363                         return;
364                 
365                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
366
367                 /* max freq cannot be less than 100. But who knows.... */
368                 if (unlikely(freq_step == 0))
369                         freq_step = 5;
370                 
371                 this_dbs_info->requested_freq += freq_step;
372                 if (this_dbs_info->requested_freq > policy->max)
373                         this_dbs_info->requested_freq = policy->max;
374
375                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
376                         CPUFREQ_RELATION_H);
377                 return;
378         }
379
380         /* Check for frequency decrease */
381         this_dbs_info->down_skip++;
382         if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
383                 return;
384
385         /* Check for frequency decrease */
386         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
387         tmp_idle_ticks = total_idle_ticks -
388                 this_dbs_info->prev_cpu_idle_down;
389         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
390
391         if (tmp_idle_ticks < idle_ticks)
392                 idle_ticks = tmp_idle_ticks;
393
394         /* Scale idle ticks by 100 and compare with up and down ticks */
395         idle_ticks *= 100;
396         this_dbs_info->down_skip = 0;
397
398         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
399                 dbs_tuners_ins.sampling_down_factor;
400         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
401                 usecs_to_jiffies(freq_down_sampling_rate);
402
403         if (idle_ticks > down_idle_ticks) {
404                 /*
405                  * if we are already at the lowest speed then break out early
406                  * or if we 'cannot' reduce the speed as the user might want
407                  * freq_step to be zero
408                  */
409                 if (this_dbs_info->requested_freq == policy->min
410                                 || dbs_tuners_ins.freq_step == 0)
411                         return;
412
413                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
414
415                 /* max freq cannot be less than 100. But who knows.... */
416                 if (unlikely(freq_step == 0))
417                         freq_step = 5;
418
419                 this_dbs_info->requested_freq -= freq_step;
420                 if (this_dbs_info->requested_freq < policy->min)
421                         this_dbs_info->requested_freq = policy->min;
422
423                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
424                                 CPUFREQ_RELATION_H);
425                 return;
426         }
427 }
428
429 static void do_dbs_timer(struct work_struct *work)
430
431         int i;
432         mutex_lock(&dbs_mutex);
433         for_each_online_cpu(i)
434                 dbs_check_cpu(i);
435         schedule_delayed_work(&dbs_work, 
436                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
437         mutex_unlock(&dbs_mutex);
438
439
440 static inline void dbs_timer_init(void)
441 {
442         schedule_delayed_work(&dbs_work,
443                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
444         return;
445 }
446
447 static inline void dbs_timer_exit(void)
448 {
449         cancel_delayed_work(&dbs_work);
450         return;
451 }
452
453 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
454                                    unsigned int event)
455 {
456         unsigned int cpu = policy->cpu;
457         struct cpu_dbs_info_s *this_dbs_info;
458         unsigned int j;
459         int rc;
460
461         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
462
463         switch (event) {
464         case CPUFREQ_GOV_START:
465                 if ((!cpu_online(cpu)) || 
466                     (!policy->cur))
467                         return -EINVAL;
468
469                 if (policy->cpuinfo.transition_latency >
470                                 (TRANSITION_LATENCY_LIMIT * 1000))
471                         return -EINVAL;
472                 if (this_dbs_info->enable) /* Already enabled */
473                         break;
474                  
475                 mutex_lock(&dbs_mutex);
476
477                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
478                 if (rc) {
479                         mutex_unlock(&dbs_mutex);
480                         return rc;
481                 }
482
483                 for_each_cpu_mask(j, policy->cpus) {
484                         struct cpu_dbs_info_s *j_dbs_info;
485                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
486                         j_dbs_info->cur_policy = policy;
487                 
488                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
489                         j_dbs_info->prev_cpu_idle_down
490                                 = j_dbs_info->prev_cpu_idle_up;
491                 }
492                 this_dbs_info->enable = 1;
493                 this_dbs_info->down_skip = 0;
494                 this_dbs_info->requested_freq = policy->cur;
495
496                 dbs_enable++;
497                 /*
498                  * Start the timerschedule work, when this governor
499                  * is used for first time
500                  */
501                 if (dbs_enable == 1) {
502                         unsigned int latency;
503                         /* policy latency is in nS. Convert it to uS first */
504                         latency = policy->cpuinfo.transition_latency / 1000;
505                         if (latency == 0)
506                                 latency = 1;
507
508                         def_sampling_rate = 10 * latency *
509                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
510
511                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
512                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
513
514                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
515
516                         dbs_timer_init();
517                 }
518                 
519                 mutex_unlock(&dbs_mutex);
520                 break;
521
522         case CPUFREQ_GOV_STOP:
523                 mutex_lock(&dbs_mutex);
524                 this_dbs_info->enable = 0;
525                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
526                 dbs_enable--;
527                 /*
528                  * Stop the timerschedule work, when this governor
529                  * is used for first time
530                  */
531                 if (dbs_enable == 0) 
532                         dbs_timer_exit();
533                 
534                 mutex_unlock(&dbs_mutex);
535
536                 break;
537
538         case CPUFREQ_GOV_LIMITS:
539                 mutex_lock(&dbs_mutex);
540                 if (policy->max < this_dbs_info->cur_policy->cur)
541                         __cpufreq_driver_target(
542                                         this_dbs_info->cur_policy,
543                                         policy->max, CPUFREQ_RELATION_H);
544                 else if (policy->min > this_dbs_info->cur_policy->cur)
545                         __cpufreq_driver_target(
546                                         this_dbs_info->cur_policy,
547                                         policy->min, CPUFREQ_RELATION_L);
548                 mutex_unlock(&dbs_mutex);
549                 break;
550         }
551         return 0;
552 }
553
554 static struct cpufreq_governor cpufreq_gov_dbs = {
555         .name           = "conservative",
556         .governor       = cpufreq_governor_dbs,
557         .owner          = THIS_MODULE,
558 };
559
560 static int __init cpufreq_gov_dbs_init(void)
561 {
562         return cpufreq_register_governor(&cpufreq_gov_dbs);
563 }
564
565 static void __exit cpufreq_gov_dbs_exit(void)
566 {
567         /* Make sure that the scheduled work is indeed not running */
568         flush_scheduled_work();
569
570         cpufreq_unregister_governor(&cpufreq_gov_dbs);
571 }
572
573
574 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
575 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
576                 "Low Latency Frequency Transition capable processors "
577                 "optimised for use in a battery environment");
578 MODULE_LICENSE ("GPL");
579
580 module_init(cpufreq_gov_dbs_init);
581 module_exit(cpufreq_gov_dbs_exit);