[CPUFREQ] Remove old, deprecated per cpu ondemand/conservative sysfs files
[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)  2009 Alexander Clouter <alex@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/init.h>
17 #include <linux/cpufreq.h>
18 #include <linux/cpu.h>
19 #include <linux/jiffies.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/mutex.h>
22 #include <linux/hrtimer.h>
23 #include <linux/tick.h>
24 #include <linux/ktime.h>
25 #include <linux/sched.h>
26
27 /*
28  * dbs is used in this file as a shortform for demandbased switching
29  * It helps to keep variable names smaller, simpler
30  */
31
32 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
33 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
34
35 /*
36  * The polling frequency of this governor depends on the capability of
37  * the processor. Default polling frequency is 1000 times the transition
38  * latency of the processor. The governor will work on any processor with
39  * transition latency <= 10mS, using appropriate sampling
40  * rate.
41  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
42  * this governor will not work.
43  * All times here are in uS.
44  */
45 #define MIN_SAMPLING_RATE_RATIO                 (2)
46
47 static unsigned int min_sampling_rate;
48
49 #define LATENCY_MULTIPLIER                      (1000)
50 #define MIN_LATENCY_MULTIPLIER                  (100)
51 #define DEF_SAMPLING_DOWN_FACTOR                (1)
52 #define MAX_SAMPLING_DOWN_FACTOR                (10)
53 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
54
55 static void do_dbs_timer(struct work_struct *work);
56
57 struct cpu_dbs_info_s {
58         cputime64_t prev_cpu_idle;
59         cputime64_t prev_cpu_wall;
60         cputime64_t prev_cpu_nice;
61         struct cpufreq_policy *cur_policy;
62         struct delayed_work work;
63         unsigned int down_skip;
64         unsigned int requested_freq;
65         int cpu;
66         unsigned int enable:1;
67         /*
68          * percpu mutex that serializes governor limit change with
69          * do_dbs_timer invocation. We do not want do_dbs_timer to run
70          * when user is changing the governor or limits.
71          */
72         struct mutex timer_mutex;
73 };
74 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info);
75
76 static unsigned int dbs_enable; /* number of CPUs using this policy */
77
78 /*
79  * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
80  * different CPUs. It protects dbs_enable in governor start/stop.
81  */
82 static DEFINE_MUTEX(dbs_mutex);
83
84 static struct dbs_tuners {
85         unsigned int sampling_rate;
86         unsigned int sampling_down_factor;
87         unsigned int up_threshold;
88         unsigned int down_threshold;
89         unsigned int ignore_nice;
90         unsigned int freq_step;
91 } dbs_tuners_ins = {
92         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
93         .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
94         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
95         .ignore_nice = 0,
96         .freq_step = 5,
97 };
98
99 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
100                                                         cputime64_t *wall)
101 {
102         cputime64_t idle_time;
103         cputime64_t cur_wall_time;
104         cputime64_t busy_time;
105
106         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
107         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
108                         kstat_cpu(cpu).cpustat.system);
109
110         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
111         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
112         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
113         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
114
115         idle_time = cputime64_sub(cur_wall_time, busy_time);
116         if (wall)
117                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
118
119         return (cputime64_t)jiffies_to_usecs(idle_time);
120 }
121
122 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
123 {
124         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
125
126         if (idle_time == -1ULL)
127                 return get_cpu_idle_time_jiffy(cpu, wall);
128
129         return idle_time;
130 }
131
132 /* keep track of frequency transitions */
133 static int
134 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
135                      void *data)
136 {
137         struct cpufreq_freqs *freq = data;
138         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info,
139                                                         freq->cpu);
140
141         struct cpufreq_policy *policy;
142
143         if (!this_dbs_info->enable)
144                 return 0;
145
146         policy = this_dbs_info->cur_policy;
147
148         /*
149          * we only care if our internally tracked freq moves outside
150          * the 'valid' ranges of freqency available to us otherwise
151          * we do not change it
152         */
153         if (this_dbs_info->requested_freq > policy->max
154                         || this_dbs_info->requested_freq < policy->min)
155                 this_dbs_info->requested_freq = freq->new;
156
157         return 0;
158 }
159
160 static struct notifier_block dbs_cpufreq_notifier_block = {
161         .notifier_call = dbs_cpufreq_notifier
162 };
163
164 /************************** sysfs interface ************************/
165 static ssize_t show_sampling_rate_min(struct kobject *kobj,
166                                       struct attribute *attr, char *buf)
167 {
168         return sprintf(buf, "%u\n", min_sampling_rate);
169 }
170
171 define_one_global_ro(sampling_rate_min);
172
173 /* cpufreq_conservative Governor Tunables */
174 #define show_one(file_name, object)                                     \
175 static ssize_t show_##file_name                                         \
176 (struct kobject *kobj, struct attribute *attr, char *buf)               \
177 {                                                                       \
178         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
179 }
180 show_one(sampling_rate, sampling_rate);
181 show_one(sampling_down_factor, sampling_down_factor);
182 show_one(up_threshold, up_threshold);
183 show_one(down_threshold, down_threshold);
184 show_one(ignore_nice_load, ignore_nice);
185 show_one(freq_step, freq_step);
186
187 static ssize_t store_sampling_down_factor(struct kobject *a,
188                                           struct attribute *b,
189                                           const char *buf, size_t count)
190 {
191         unsigned int input;
192         int ret;
193         ret = sscanf(buf, "%u", &input);
194
195         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
196                 return -EINVAL;
197
198         mutex_lock(&dbs_mutex);
199         dbs_tuners_ins.sampling_down_factor = input;
200         mutex_unlock(&dbs_mutex);
201
202         return count;
203 }
204
205 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
206                                    const char *buf, size_t count)
207 {
208         unsigned int input;
209         int ret;
210         ret = sscanf(buf, "%u", &input);
211
212         if (ret != 1)
213                 return -EINVAL;
214
215         mutex_lock(&dbs_mutex);
216         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
217         mutex_unlock(&dbs_mutex);
218
219         return count;
220 }
221
222 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
223                                   const char *buf, size_t count)
224 {
225         unsigned int input;
226         int ret;
227         ret = sscanf(buf, "%u", &input);
228
229         mutex_lock(&dbs_mutex);
230         if (ret != 1 || input > 100 ||
231                         input <= dbs_tuners_ins.down_threshold) {
232                 mutex_unlock(&dbs_mutex);
233                 return -EINVAL;
234         }
235
236         dbs_tuners_ins.up_threshold = input;
237         mutex_unlock(&dbs_mutex);
238
239         return count;
240 }
241
242 static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
243                                     const char *buf, size_t count)
244 {
245         unsigned int input;
246         int ret;
247         ret = sscanf(buf, "%u", &input);
248
249         mutex_lock(&dbs_mutex);
250         /* cannot be lower than 11 otherwise freq will not fall */
251         if (ret != 1 || input < 11 || input > 100 ||
252                         input >= dbs_tuners_ins.up_threshold) {
253                 mutex_unlock(&dbs_mutex);
254                 return -EINVAL;
255         }
256
257         dbs_tuners_ins.down_threshold = input;
258         mutex_unlock(&dbs_mutex);
259
260         return count;
261 }
262
263 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
264                                       const char *buf, size_t count)
265 {
266         unsigned int input;
267         int ret;
268
269         unsigned int j;
270
271         ret = sscanf(buf, "%u", &input);
272         if (ret != 1)
273                 return -EINVAL;
274
275         if (input > 1)
276                 input = 1;
277
278         mutex_lock(&dbs_mutex);
279         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
280                 mutex_unlock(&dbs_mutex);
281                 return count;
282         }
283         dbs_tuners_ins.ignore_nice = input;
284
285         /* we need to re-evaluate prev_cpu_idle */
286         for_each_online_cpu(j) {
287                 struct cpu_dbs_info_s *dbs_info;
288                 dbs_info = &per_cpu(cs_cpu_dbs_info, j);
289                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
290                                                 &dbs_info->prev_cpu_wall);
291                 if (dbs_tuners_ins.ignore_nice)
292                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
293         }
294         mutex_unlock(&dbs_mutex);
295
296         return count;
297 }
298
299 static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
300                                const char *buf, size_t count)
301 {
302         unsigned int input;
303         int ret;
304         ret = sscanf(buf, "%u", &input);
305
306         if (ret != 1)
307                 return -EINVAL;
308
309         if (input > 100)
310                 input = 100;
311
312         /* no need to test here if freq_step is zero as the user might actually
313          * want this, they would be crazy though :) */
314         mutex_lock(&dbs_mutex);
315         dbs_tuners_ins.freq_step = input;
316         mutex_unlock(&dbs_mutex);
317
318         return count;
319 }
320
321 define_one_global_rw(sampling_rate);
322 define_one_global_rw(sampling_down_factor);
323 define_one_global_rw(up_threshold);
324 define_one_global_rw(down_threshold);
325 define_one_global_rw(ignore_nice_load);
326 define_one_global_rw(freq_step);
327
328 static struct attribute *dbs_attributes[] = {
329         &sampling_rate_min.attr,
330         &sampling_rate.attr,
331         &sampling_down_factor.attr,
332         &up_threshold.attr,
333         &down_threshold.attr,
334         &ignore_nice_load.attr,
335         &freq_step.attr,
336         NULL
337 };
338
339 static struct attribute_group dbs_attr_group = {
340         .attrs = dbs_attributes,
341         .name = "conservative",
342 };
343
344 /************************** sysfs end ************************/
345
346 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
347 {
348         unsigned int load = 0;
349         unsigned int max_load = 0;
350         unsigned int freq_target;
351
352         struct cpufreq_policy *policy;
353         unsigned int j;
354
355         policy = this_dbs_info->cur_policy;
356
357         /*
358          * Every sampling_rate, we check, if current idle time is less
359          * than 20% (default), then we try to increase frequency
360          * Every sampling_rate*sampling_down_factor, we check, if current
361          * idle time is more than 80%, then we try to decrease frequency
362          *
363          * Any frequency increase takes it to the maximum frequency.
364          * Frequency reduction happens at minimum steps of
365          * 5% (default) of maximum frequency
366          */
367
368         /* Get Absolute Load */
369         for_each_cpu(j, policy->cpus) {
370                 struct cpu_dbs_info_s *j_dbs_info;
371                 cputime64_t cur_wall_time, cur_idle_time;
372                 unsigned int idle_time, wall_time;
373
374                 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
375
376                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
377
378                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
379                                 j_dbs_info->prev_cpu_wall);
380                 j_dbs_info->prev_cpu_wall = cur_wall_time;
381
382                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
383                                 j_dbs_info->prev_cpu_idle);
384                 j_dbs_info->prev_cpu_idle = cur_idle_time;
385
386                 if (dbs_tuners_ins.ignore_nice) {
387                         cputime64_t cur_nice;
388                         unsigned long cur_nice_jiffies;
389
390                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
391                                          j_dbs_info->prev_cpu_nice);
392                         /*
393                          * Assumption: nice time between sampling periods will
394                          * be less than 2^32 jiffies for 32 bit sys
395                          */
396                         cur_nice_jiffies = (unsigned long)
397                                         cputime64_to_jiffies64(cur_nice);
398
399                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
400                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
401                 }
402
403                 if (unlikely(!wall_time || wall_time < idle_time))
404                         continue;
405
406                 load = 100 * (wall_time - idle_time) / wall_time;
407
408                 if (load > max_load)
409                         max_load = load;
410         }
411
412         /*
413          * break out if we 'cannot' reduce the speed as the user might
414          * want freq_step to be zero
415          */
416         if (dbs_tuners_ins.freq_step == 0)
417                 return;
418
419         /* Check for frequency increase */
420         if (max_load > dbs_tuners_ins.up_threshold) {
421                 this_dbs_info->down_skip = 0;
422
423                 /* if we are already at full speed then break out early */
424                 if (this_dbs_info->requested_freq == policy->max)
425                         return;
426
427                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
428
429                 /* max freq cannot be less than 100. But who knows.... */
430                 if (unlikely(freq_target == 0))
431                         freq_target = 5;
432
433                 this_dbs_info->requested_freq += freq_target;
434                 if (this_dbs_info->requested_freq > policy->max)
435                         this_dbs_info->requested_freq = policy->max;
436
437                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
438                         CPUFREQ_RELATION_H);
439                 return;
440         }
441
442         /*
443          * The optimal frequency is the frequency that is the lowest that
444          * can support the current CPU usage without triggering the up
445          * policy. To be safe, we focus 10 points under the threshold.
446          */
447         if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
448                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
449
450                 this_dbs_info->requested_freq -= freq_target;
451                 if (this_dbs_info->requested_freq < policy->min)
452                         this_dbs_info->requested_freq = policy->min;
453
454                 /*
455                  * if we cannot reduce the frequency anymore, break out early
456                  */
457                 if (policy->cur == policy->min)
458                         return;
459
460                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
461                                 CPUFREQ_RELATION_H);
462                 return;
463         }
464 }
465
466 static void do_dbs_timer(struct work_struct *work)
467 {
468         struct cpu_dbs_info_s *dbs_info =
469                 container_of(work, struct cpu_dbs_info_s, work.work);
470         unsigned int cpu = dbs_info->cpu;
471
472         /* We want all CPUs to do sampling nearly on same jiffy */
473         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
474
475         delay -= jiffies % delay;
476
477         mutex_lock(&dbs_info->timer_mutex);
478
479         dbs_check_cpu(dbs_info);
480
481         schedule_delayed_work_on(cpu, &dbs_info->work, delay);
482         mutex_unlock(&dbs_info->timer_mutex);
483 }
484
485 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
486 {
487         /* We want all CPUs to do sampling nearly on same jiffy */
488         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
489         delay -= jiffies % delay;
490
491         dbs_info->enable = 1;
492         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
493         schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
494 }
495
496 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
497 {
498         dbs_info->enable = 0;
499         cancel_delayed_work_sync(&dbs_info->work);
500 }
501
502 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
503                                    unsigned int event)
504 {
505         unsigned int cpu = policy->cpu;
506         struct cpu_dbs_info_s *this_dbs_info;
507         unsigned int j;
508         int rc;
509
510         this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
511
512         switch (event) {
513         case CPUFREQ_GOV_START:
514                 if ((!cpu_online(cpu)) || (!policy->cur))
515                         return -EINVAL;
516
517                 mutex_lock(&dbs_mutex);
518
519                 for_each_cpu(j, policy->cpus) {
520                         struct cpu_dbs_info_s *j_dbs_info;
521                         j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
522                         j_dbs_info->cur_policy = policy;
523
524                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
525                                                 &j_dbs_info->prev_cpu_wall);
526                         if (dbs_tuners_ins.ignore_nice) {
527                                 j_dbs_info->prev_cpu_nice =
528                                                 kstat_cpu(j).cpustat.nice;
529                         }
530                 }
531                 this_dbs_info->down_skip = 0;
532                 this_dbs_info->requested_freq = policy->cur;
533
534                 mutex_init(&this_dbs_info->timer_mutex);
535                 dbs_enable++;
536                 /*
537                  * Start the timerschedule work, when this governor
538                  * is used for first time
539                  */
540                 if (dbs_enable == 1) {
541                         unsigned int latency;
542                         /* policy latency is in nS. Convert it to uS first */
543                         latency = policy->cpuinfo.transition_latency / 1000;
544                         if (latency == 0)
545                                 latency = 1;
546
547                         rc = sysfs_create_group(cpufreq_global_kobject,
548                                                 &dbs_attr_group);
549                         if (rc) {
550                                 mutex_unlock(&dbs_mutex);
551                                 return rc;
552                         }
553
554                         /*
555                          * conservative does not implement micro like ondemand
556                          * governor, thus we are bound to jiffes/HZ
557                          */
558                         min_sampling_rate =
559                                 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
560                         /* Bring kernel and HW constraints together */
561                         min_sampling_rate = max(min_sampling_rate,
562                                         MIN_LATENCY_MULTIPLIER * latency);
563                         dbs_tuners_ins.sampling_rate =
564                                 max(min_sampling_rate,
565                                     latency * LATENCY_MULTIPLIER);
566
567                         cpufreq_register_notifier(
568                                         &dbs_cpufreq_notifier_block,
569                                         CPUFREQ_TRANSITION_NOTIFIER);
570                 }
571                 mutex_unlock(&dbs_mutex);
572
573                 dbs_timer_init(this_dbs_info);
574
575                 break;
576
577         case CPUFREQ_GOV_STOP:
578                 dbs_timer_exit(this_dbs_info);
579
580                 mutex_lock(&dbs_mutex);
581                 dbs_enable--;
582                 mutex_destroy(&this_dbs_info->timer_mutex);
583
584                 /*
585                  * Stop the timerschedule work, when this governor
586                  * is used for first time
587                  */
588                 if (dbs_enable == 0)
589                         cpufreq_unregister_notifier(
590                                         &dbs_cpufreq_notifier_block,
591                                         CPUFREQ_TRANSITION_NOTIFIER);
592
593                 mutex_unlock(&dbs_mutex);
594                 if (!dbs_enable)
595                         sysfs_remove_group(cpufreq_global_kobject,
596                                            &dbs_attr_group);
597
598                 break;
599
600         case CPUFREQ_GOV_LIMITS:
601                 mutex_lock(&this_dbs_info->timer_mutex);
602                 if (policy->max < this_dbs_info->cur_policy->cur)
603                         __cpufreq_driver_target(
604                                         this_dbs_info->cur_policy,
605                                         policy->max, CPUFREQ_RELATION_H);
606                 else if (policy->min > this_dbs_info->cur_policy->cur)
607                         __cpufreq_driver_target(
608                                         this_dbs_info->cur_policy,
609                                         policy->min, CPUFREQ_RELATION_L);
610                 mutex_unlock(&this_dbs_info->timer_mutex);
611
612                 break;
613         }
614         return 0;
615 }
616
617 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
618 static
619 #endif
620 struct cpufreq_governor cpufreq_gov_conservative = {
621         .name                   = "conservative",
622         .governor               = cpufreq_governor_dbs,
623         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
624         .owner                  = THIS_MODULE,
625 };
626
627 static int __init cpufreq_gov_dbs_init(void)
628 {
629         return cpufreq_register_governor(&cpufreq_gov_conservative);
630 }
631
632 static void __exit cpufreq_gov_dbs_exit(void)
633 {
634         cpufreq_unregister_governor(&cpufreq_gov_conservative);
635 }
636
637
638 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
639 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
640                 "Low Latency Frequency Transition capable processors "
641                 "optimised for use in a battery environment");
642 MODULE_LICENSE("GPL");
643
644 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
645 fs_initcall(cpufreq_gov_dbs_init);
646 #else
647 module_init(cpufreq_gov_dbs_init);
648 #endif
649 module_exit(cpufreq_gov_dbs_exit);