Merge branch 'release-2.6.27' of git://git.kernel.org/pub/scm/linux/kernel/git/ak...
[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/init.h>
16 #include <linux/cpufreq.h>
17 #include <linux/cpu.h>
18 #include <linux/jiffies.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mutex.h>
21
22 /*
23  * dbs is used in this file as a shortform for demandbased switching
24  * It helps to keep variable names smaller, simpler
25  */
26
27 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
28 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
29 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
30
31 /*
32  * The polling frequency of this governor depends on the capability of
33  * the processor. Default polling frequency is 1000 times the transition
34  * latency of the processor. The governor will work on any processor with
35  * transition latency <= 10mS, using appropriate sampling
36  * rate.
37  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
38  * this governor will not work.
39  * All times here are in uS.
40  */
41 static unsigned int def_sampling_rate;
42 #define MIN_SAMPLING_RATE_RATIO                 (2)
43 /* for correct statistics, we need at least 10 ticks between each measure */
44 #define MIN_STAT_SAMPLING_RATE                  \
45                         (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
46 #define MIN_SAMPLING_RATE                       \
47                         (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
48 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
49 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
50 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
51
52 static void do_dbs_timer(struct work_struct *work);
53
54 /* Sampling types */
55 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
56
57 struct cpu_dbs_info_s {
58         cputime64_t prev_cpu_idle;
59         cputime64_t prev_cpu_wall;
60         struct cpufreq_policy *cur_policy;
61         struct delayed_work work;
62         struct cpufreq_frequency_table *freq_table;
63         unsigned int freq_lo;
64         unsigned int freq_lo_jiffies;
65         unsigned int freq_hi_jiffies;
66         int cpu;
67         unsigned int enable:1,
68                      sample_type:1;
69 };
70 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
71
72 static unsigned int dbs_enable; /* number of CPUs using this policy */
73
74 /*
75  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
76  * lock and dbs_mutex. cpu_hotplug lock should always be held before
77  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
78  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
79  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
80  * is recursive for the same process. -Venki
81  */
82 static DEFINE_MUTEX(dbs_mutex);
83
84 static struct workqueue_struct  *kondemand_wq;
85
86 static struct dbs_tuners {
87         unsigned int sampling_rate;
88         unsigned int up_threshold;
89         unsigned int ignore_nice;
90         unsigned int powersave_bias;
91 } dbs_tuners_ins = {
92         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
93         .ignore_nice = 0,
94         .powersave_bias = 0,
95 };
96
97 static inline cputime64_t get_cpu_idle_time(unsigned int cpu)
98 {
99         cputime64_t idle_time;
100         cputime64_t cur_jiffies;
101         cputime64_t busy_time;
102
103         cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
104         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
105                         kstat_cpu(cpu).cpustat.system);
106
107         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
108         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
109         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
110
111         if (!dbs_tuners_ins.ignore_nice) {
112                 busy_time = cputime64_add(busy_time,
113                                 kstat_cpu(cpu).cpustat.nice);
114         }
115
116         idle_time = cputime64_sub(cur_jiffies, busy_time);
117         return idle_time;
118 }
119
120 /*
121  * Find right freq to be set now with powersave_bias on.
122  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
123  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
124  */
125 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
126                                           unsigned int freq_next,
127                                           unsigned int relation)
128 {
129         unsigned int freq_req, freq_reduc, freq_avg;
130         unsigned int freq_hi, freq_lo;
131         unsigned int index = 0;
132         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
133         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
134
135         if (!dbs_info->freq_table) {
136                 dbs_info->freq_lo = 0;
137                 dbs_info->freq_lo_jiffies = 0;
138                 return freq_next;
139         }
140
141         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
142                         relation, &index);
143         freq_req = dbs_info->freq_table[index].frequency;
144         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
145         freq_avg = freq_req - freq_reduc;
146
147         /* Find freq bounds for freq_avg in freq_table */
148         index = 0;
149         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
150                         CPUFREQ_RELATION_H, &index);
151         freq_lo = dbs_info->freq_table[index].frequency;
152         index = 0;
153         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
154                         CPUFREQ_RELATION_L, &index);
155         freq_hi = dbs_info->freq_table[index].frequency;
156
157         /* Find out how long we have to be in hi and lo freqs */
158         if (freq_hi == freq_lo) {
159                 dbs_info->freq_lo = 0;
160                 dbs_info->freq_lo_jiffies = 0;
161                 return freq_lo;
162         }
163         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
164         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
165         jiffies_hi += ((freq_hi - freq_lo) / 2);
166         jiffies_hi /= (freq_hi - freq_lo);
167         jiffies_lo = jiffies_total - jiffies_hi;
168         dbs_info->freq_lo = freq_lo;
169         dbs_info->freq_lo_jiffies = jiffies_lo;
170         dbs_info->freq_hi_jiffies = jiffies_hi;
171         return freq_hi;
172 }
173
174 static void ondemand_powersave_bias_init(void)
175 {
176         int i;
177         for_each_online_cpu(i) {
178                 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
179                 dbs_info->freq_table = cpufreq_frequency_get_table(i);
180                 dbs_info->freq_lo = 0;
181         }
182 }
183
184 /************************** sysfs interface ************************/
185 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
186 {
187         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
188 }
189
190 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
191 {
192         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
193 }
194
195 #define define_one_ro(_name)            \
196 static struct freq_attr _name =         \
197 __ATTR(_name, 0444, show_##_name, NULL)
198
199 define_one_ro(sampling_rate_max);
200 define_one_ro(sampling_rate_min);
201
202 /* cpufreq_ondemand Governor Tunables */
203 #define show_one(file_name, object)                                     \
204 static ssize_t show_##file_name                                         \
205 (struct cpufreq_policy *unused, char *buf)                              \
206 {                                                                       \
207         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
208 }
209 show_one(sampling_rate, sampling_rate);
210 show_one(up_threshold, up_threshold);
211 show_one(ignore_nice_load, ignore_nice);
212 show_one(powersave_bias, powersave_bias);
213
214 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
215                 const char *buf, size_t count)
216 {
217         unsigned int input;
218         int ret;
219         ret = sscanf(buf, "%u", &input);
220
221         mutex_lock(&dbs_mutex);
222         if (ret != 1 || input > MAX_SAMPLING_RATE
223                      || input < MIN_SAMPLING_RATE) {
224                 mutex_unlock(&dbs_mutex);
225                 return -EINVAL;
226         }
227
228         dbs_tuners_ins.sampling_rate = input;
229         mutex_unlock(&dbs_mutex);
230
231         return count;
232 }
233
234 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
235                 const char *buf, size_t count)
236 {
237         unsigned int input;
238         int ret;
239         ret = sscanf(buf, "%u", &input);
240
241         mutex_lock(&dbs_mutex);
242         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
243                         input < MIN_FREQUENCY_UP_THRESHOLD) {
244                 mutex_unlock(&dbs_mutex);
245                 return -EINVAL;
246         }
247
248         dbs_tuners_ins.up_threshold = input;
249         mutex_unlock(&dbs_mutex);
250
251         return count;
252 }
253
254 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
255                 const char *buf, size_t count)
256 {
257         unsigned int input;
258         int ret;
259
260         unsigned int j;
261
262         ret = sscanf(buf, "%u", &input);
263         if ( ret != 1 )
264                 return -EINVAL;
265
266         if ( input > 1 )
267                 input = 1;
268
269         mutex_lock(&dbs_mutex);
270         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
271                 mutex_unlock(&dbs_mutex);
272                 return count;
273         }
274         dbs_tuners_ins.ignore_nice = input;
275
276         /* we need to re-evaluate prev_cpu_idle */
277         for_each_online_cpu(j) {
278                 struct cpu_dbs_info_s *dbs_info;
279                 dbs_info = &per_cpu(cpu_dbs_info, j);
280                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
281                 dbs_info->prev_cpu_wall = get_jiffies_64();
282         }
283         mutex_unlock(&dbs_mutex);
284
285         return count;
286 }
287
288 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
289                 const char *buf, size_t count)
290 {
291         unsigned int input;
292         int ret;
293         ret = sscanf(buf, "%u", &input);
294
295         if (ret != 1)
296                 return -EINVAL;
297
298         if (input > 1000)
299                 input = 1000;
300
301         mutex_lock(&dbs_mutex);
302         dbs_tuners_ins.powersave_bias = input;
303         ondemand_powersave_bias_init();
304         mutex_unlock(&dbs_mutex);
305
306         return count;
307 }
308
309 #define define_one_rw(_name) \
310 static struct freq_attr _name = \
311 __ATTR(_name, 0644, show_##_name, store_##_name)
312
313 define_one_rw(sampling_rate);
314 define_one_rw(up_threshold);
315 define_one_rw(ignore_nice_load);
316 define_one_rw(powersave_bias);
317
318 static struct attribute * dbs_attributes[] = {
319         &sampling_rate_max.attr,
320         &sampling_rate_min.attr,
321         &sampling_rate.attr,
322         &up_threshold.attr,
323         &ignore_nice_load.attr,
324         &powersave_bias.attr,
325         NULL
326 };
327
328 static struct attribute_group dbs_attr_group = {
329         .attrs = dbs_attributes,
330         .name = "ondemand",
331 };
332
333 /************************** sysfs end ************************/
334
335 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
336 {
337         unsigned int idle_ticks, total_ticks;
338         unsigned int load = 0;
339         cputime64_t cur_jiffies;
340
341         struct cpufreq_policy *policy;
342         unsigned int j;
343
344         if (!this_dbs_info->enable)
345                 return;
346
347         this_dbs_info->freq_lo = 0;
348         policy = this_dbs_info->cur_policy;
349         cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
350         total_ticks = (unsigned int) cputime64_sub(cur_jiffies,
351                         this_dbs_info->prev_cpu_wall);
352         this_dbs_info->prev_cpu_wall = get_jiffies_64();
353
354         if (!total_ticks)
355                 return;
356         /*
357          * Every sampling_rate, we check, if current idle time is less
358          * than 20% (default), then we try to increase frequency
359          * Every sampling_rate, we look for a the lowest
360          * frequency which can sustain the load while keeping idle time over
361          * 30%. If such a frequency exist, we try to decrease to this frequency.
362          *
363          * Any frequency increase takes it to the maximum frequency.
364          * Frequency reduction happens at minimum steps of
365          * 5% (default) of current frequency
366          */
367
368         /* Get Idle Time */
369         idle_ticks = UINT_MAX;
370         for_each_cpu_mask_nr(j, policy->cpus) {
371                 cputime64_t total_idle_ticks;
372                 unsigned int tmp_idle_ticks;
373                 struct cpu_dbs_info_s *j_dbs_info;
374
375                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
376                 total_idle_ticks = get_cpu_idle_time(j);
377                 tmp_idle_ticks = (unsigned int) cputime64_sub(total_idle_ticks,
378                                 j_dbs_info->prev_cpu_idle);
379                 j_dbs_info->prev_cpu_idle = total_idle_ticks;
380
381                 if (tmp_idle_ticks < idle_ticks)
382                         idle_ticks = tmp_idle_ticks;
383         }
384         if (likely(total_ticks > idle_ticks))
385                 load = (100 * (total_ticks - idle_ticks)) / total_ticks;
386
387         /* Check for frequency increase */
388         if (load > dbs_tuners_ins.up_threshold) {
389                 /* if we are already at full speed then break out early */
390                 if (!dbs_tuners_ins.powersave_bias) {
391                         if (policy->cur == policy->max)
392                                 return;
393
394                         __cpufreq_driver_target(policy, policy->max,
395                                 CPUFREQ_RELATION_H);
396                 } else {
397                         int freq = powersave_bias_target(policy, policy->max,
398                                         CPUFREQ_RELATION_H);
399                         __cpufreq_driver_target(policy, freq,
400                                 CPUFREQ_RELATION_L);
401                 }
402                 return;
403         }
404
405         /* Check for frequency decrease */
406         /* if we cannot reduce the frequency anymore, break out early */
407         if (policy->cur == policy->min)
408                 return;
409
410         /*
411          * The optimal frequency is the frequency that is the lowest that
412          * can support the current CPU usage without triggering the up
413          * policy. To be safe, we focus 10 points under the threshold.
414          */
415         if (load < (dbs_tuners_ins.up_threshold - 10)) {
416                 unsigned int freq_next, freq_cur;
417
418                 freq_cur = __cpufreq_driver_getavg(policy);
419                 if (!freq_cur)
420                         freq_cur = policy->cur;
421
422                 freq_next = (freq_cur * load) /
423                         (dbs_tuners_ins.up_threshold - 10);
424
425                 if (!dbs_tuners_ins.powersave_bias) {
426                         __cpufreq_driver_target(policy, freq_next,
427                                         CPUFREQ_RELATION_L);
428                 } else {
429                         int freq = powersave_bias_target(policy, freq_next,
430                                         CPUFREQ_RELATION_L);
431                         __cpufreq_driver_target(policy, freq,
432                                 CPUFREQ_RELATION_L);
433                 }
434         }
435 }
436
437 static void do_dbs_timer(struct work_struct *work)
438 {
439         struct cpu_dbs_info_s *dbs_info =
440                 container_of(work, struct cpu_dbs_info_s, work.work);
441         unsigned int cpu = dbs_info->cpu;
442         int sample_type = dbs_info->sample_type;
443
444         /* We want all CPUs to do sampling nearly on same jiffy */
445         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
446
447         delay -= jiffies % delay;
448
449         if (lock_policy_rwsem_write(cpu) < 0)
450                 return;
451
452         if (!dbs_info->enable) {
453                 unlock_policy_rwsem_write(cpu);
454                 return;
455         }
456
457         /* Common NORMAL_SAMPLE setup */
458         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
459         if (!dbs_tuners_ins.powersave_bias ||
460             sample_type == DBS_NORMAL_SAMPLE) {
461                 dbs_check_cpu(dbs_info);
462                 if (dbs_info->freq_lo) {
463                         /* Setup timer for SUB_SAMPLE */
464                         dbs_info->sample_type = DBS_SUB_SAMPLE;
465                         delay = dbs_info->freq_hi_jiffies;
466                 }
467         } else {
468                 __cpufreq_driver_target(dbs_info->cur_policy,
469                                         dbs_info->freq_lo,
470                                         CPUFREQ_RELATION_H);
471         }
472         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
473         unlock_policy_rwsem_write(cpu);
474 }
475
476 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
477 {
478         /* We want all CPUs to do sampling nearly on same jiffy */
479         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
480         delay -= jiffies % delay;
481
482         dbs_info->enable = 1;
483         ondemand_powersave_bias_init();
484         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
485         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
486         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
487                               delay);
488 }
489
490 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
491 {
492         dbs_info->enable = 0;
493         cancel_delayed_work(&dbs_info->work);
494 }
495
496 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
497                                    unsigned int event)
498 {
499         unsigned int cpu = policy->cpu;
500         struct cpu_dbs_info_s *this_dbs_info;
501         unsigned int j;
502         int rc;
503
504         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
505
506         switch (event) {
507         case CPUFREQ_GOV_START:
508                 if ((!cpu_online(cpu)) || (!policy->cur))
509                         return -EINVAL;
510
511                 if (this_dbs_info->enable) /* Already enabled */
512                         break;
513
514                 mutex_lock(&dbs_mutex);
515                 dbs_enable++;
516
517                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
518                 if (rc) {
519                         dbs_enable--;
520                         mutex_unlock(&dbs_mutex);
521                         return rc;
522                 }
523
524                 for_each_cpu_mask_nr(j, policy->cpus) {
525                         struct cpu_dbs_info_s *j_dbs_info;
526                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
527                         j_dbs_info->cur_policy = policy;
528
529                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
530                         j_dbs_info->prev_cpu_wall = get_jiffies_64();
531                 }
532                 this_dbs_info->cpu = cpu;
533                 /*
534                  * Start the timerschedule work, when this governor
535                  * is used for first time
536                  */
537                 if (dbs_enable == 1) {
538                         unsigned int latency;
539                         /* policy latency is in nS. Convert it to uS first */
540                         latency = policy->cpuinfo.transition_latency / 1000;
541                         if (latency == 0)
542                                 latency = 1;
543
544                         def_sampling_rate = latency *
545                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
546
547                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
548                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
549
550                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
551                 }
552                 dbs_timer_init(this_dbs_info);
553
554                 mutex_unlock(&dbs_mutex);
555                 break;
556
557         case CPUFREQ_GOV_STOP:
558                 mutex_lock(&dbs_mutex);
559                 dbs_timer_exit(this_dbs_info);
560                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
561                 dbs_enable--;
562                 mutex_unlock(&dbs_mutex);
563
564                 break;
565
566         case CPUFREQ_GOV_LIMITS:
567                 mutex_lock(&dbs_mutex);
568                 if (policy->max < this_dbs_info->cur_policy->cur)
569                         __cpufreq_driver_target(this_dbs_info->cur_policy,
570                                                 policy->max,
571                                                 CPUFREQ_RELATION_H);
572                 else if (policy->min > this_dbs_info->cur_policy->cur)
573                         __cpufreq_driver_target(this_dbs_info->cur_policy,
574                                                 policy->min,
575                                                 CPUFREQ_RELATION_L);
576                 mutex_unlock(&dbs_mutex);
577                 break;
578         }
579         return 0;
580 }
581
582 struct cpufreq_governor cpufreq_gov_ondemand = {
583         .name                   = "ondemand",
584         .governor               = cpufreq_governor_dbs,
585         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
586         .owner                  = THIS_MODULE,
587 };
588 EXPORT_SYMBOL(cpufreq_gov_ondemand);
589
590 static int __init cpufreq_gov_dbs_init(void)
591 {
592         kondemand_wq = create_workqueue("kondemand");
593         if (!kondemand_wq) {
594                 printk(KERN_ERR "Creation of kondemand failed\n");
595                 return -EFAULT;
596         }
597         return cpufreq_register_governor(&cpufreq_gov_ondemand);
598 }
599
600 static void __exit cpufreq_gov_dbs_exit(void)
601 {
602         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
603         destroy_workqueue(kondemand_wq);
604 }
605
606
607 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
608 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
609 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
610                    "Low Latency Frequency Transition capable processors");
611 MODULE_LICENSE("GPL");
612
613 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
614 fs_initcall(cpufreq_gov_dbs_init);
615 #else
616 module_init(cpufreq_gov_dbs_init);
617 #endif
618 module_exit(cpufreq_gov_dbs_exit);