[CPUFREQ][3/6] cpufreq: get_cpu_idle_time() changes in ondemand for idle-microaccounting
[linux-2.6.git] / 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, cputime64_t *wall)
98 {
99         cputime64_t idle_time;
100         cputime64_t cur_wall_time;
101         cputime64_t busy_time;
102
103         cur_wall_time = 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_wall_time, busy_time);
117         if (wall)
118                 *wall = cur_wall_time;
119
120         return idle_time;
121 }
122
123 /*
124  * Find right freq to be set now with powersave_bias on.
125  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
126  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
127  */
128 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
129                                           unsigned int freq_next,
130                                           unsigned int relation)
131 {
132         unsigned int freq_req, freq_reduc, freq_avg;
133         unsigned int freq_hi, freq_lo;
134         unsigned int index = 0;
135         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
136         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
137
138         if (!dbs_info->freq_table) {
139                 dbs_info->freq_lo = 0;
140                 dbs_info->freq_lo_jiffies = 0;
141                 return freq_next;
142         }
143
144         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
145                         relation, &index);
146         freq_req = dbs_info->freq_table[index].frequency;
147         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
148         freq_avg = freq_req - freq_reduc;
149
150         /* Find freq bounds for freq_avg in freq_table */
151         index = 0;
152         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
153                         CPUFREQ_RELATION_H, &index);
154         freq_lo = dbs_info->freq_table[index].frequency;
155         index = 0;
156         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
157                         CPUFREQ_RELATION_L, &index);
158         freq_hi = dbs_info->freq_table[index].frequency;
159
160         /* Find out how long we have to be in hi and lo freqs */
161         if (freq_hi == freq_lo) {
162                 dbs_info->freq_lo = 0;
163                 dbs_info->freq_lo_jiffies = 0;
164                 return freq_lo;
165         }
166         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
167         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
168         jiffies_hi += ((freq_hi - freq_lo) / 2);
169         jiffies_hi /= (freq_hi - freq_lo);
170         jiffies_lo = jiffies_total - jiffies_hi;
171         dbs_info->freq_lo = freq_lo;
172         dbs_info->freq_lo_jiffies = jiffies_lo;
173         dbs_info->freq_hi_jiffies = jiffies_hi;
174         return freq_hi;
175 }
176
177 static void ondemand_powersave_bias_init(void)
178 {
179         int i;
180         for_each_online_cpu(i) {
181                 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
182                 dbs_info->freq_table = cpufreq_frequency_get_table(i);
183                 dbs_info->freq_lo = 0;
184         }
185 }
186
187 /************************** sysfs interface ************************/
188 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
189 {
190         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
191 }
192
193 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
194 {
195         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
196 }
197
198 #define define_one_ro(_name)            \
199 static struct freq_attr _name =         \
200 __ATTR(_name, 0444, show_##_name, NULL)
201
202 define_one_ro(sampling_rate_max);
203 define_one_ro(sampling_rate_min);
204
205 /* cpufreq_ondemand Governor Tunables */
206 #define show_one(file_name, object)                                     \
207 static ssize_t show_##file_name                                         \
208 (struct cpufreq_policy *unused, char *buf)                              \
209 {                                                                       \
210         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
211 }
212 show_one(sampling_rate, sampling_rate);
213 show_one(up_threshold, up_threshold);
214 show_one(ignore_nice_load, ignore_nice);
215 show_one(powersave_bias, powersave_bias);
216
217 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
218                 const char *buf, size_t count)
219 {
220         unsigned int input;
221         int ret;
222         ret = sscanf(buf, "%u", &input);
223
224         mutex_lock(&dbs_mutex);
225         if (ret != 1 || input > MAX_SAMPLING_RATE
226                      || input < MIN_SAMPLING_RATE) {
227                 mutex_unlock(&dbs_mutex);
228                 return -EINVAL;
229         }
230
231         dbs_tuners_ins.sampling_rate = input;
232         mutex_unlock(&dbs_mutex);
233
234         return count;
235 }
236
237 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
238                 const char *buf, size_t count)
239 {
240         unsigned int input;
241         int ret;
242         ret = sscanf(buf, "%u", &input);
243
244         mutex_lock(&dbs_mutex);
245         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
246                         input < MIN_FREQUENCY_UP_THRESHOLD) {
247                 mutex_unlock(&dbs_mutex);
248                 return -EINVAL;
249         }
250
251         dbs_tuners_ins.up_threshold = input;
252         mutex_unlock(&dbs_mutex);
253
254         return count;
255 }
256
257 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
258                 const char *buf, size_t count)
259 {
260         unsigned int input;
261         int ret;
262
263         unsigned int j;
264
265         ret = sscanf(buf, "%u", &input);
266         if ( ret != 1 )
267                 return -EINVAL;
268
269         if ( input > 1 )
270                 input = 1;
271
272         mutex_lock(&dbs_mutex);
273         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
274                 mutex_unlock(&dbs_mutex);
275                 return count;
276         }
277         dbs_tuners_ins.ignore_nice = input;
278
279         /* we need to re-evaluate prev_cpu_idle */
280         for_each_online_cpu(j) {
281                 struct cpu_dbs_info_s *dbs_info;
282                 dbs_info = &per_cpu(cpu_dbs_info, j);
283                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
284                                                 &dbs_info->prev_cpu_wall);
285         }
286         mutex_unlock(&dbs_mutex);
287
288         return count;
289 }
290
291 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
292                 const char *buf, size_t count)
293 {
294         unsigned int input;
295         int ret;
296         ret = sscanf(buf, "%u", &input);
297
298         if (ret != 1)
299                 return -EINVAL;
300
301         if (input > 1000)
302                 input = 1000;
303
304         mutex_lock(&dbs_mutex);
305         dbs_tuners_ins.powersave_bias = input;
306         ondemand_powersave_bias_init();
307         mutex_unlock(&dbs_mutex);
308
309         return count;
310 }
311
312 #define define_one_rw(_name) \
313 static struct freq_attr _name = \
314 __ATTR(_name, 0644, show_##_name, store_##_name)
315
316 define_one_rw(sampling_rate);
317 define_one_rw(up_threshold);
318 define_one_rw(ignore_nice_load);
319 define_one_rw(powersave_bias);
320
321 static struct attribute * dbs_attributes[] = {
322         &sampling_rate_max.attr,
323         &sampling_rate_min.attr,
324         &sampling_rate.attr,
325         &up_threshold.attr,
326         &ignore_nice_load.attr,
327         &powersave_bias.attr,
328         NULL
329 };
330
331 static struct attribute_group dbs_attr_group = {
332         .attrs = dbs_attributes,
333         .name = "ondemand",
334 };
335
336 /************************** sysfs end ************************/
337
338 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
339 {
340         unsigned int max_load_freq;
341
342         struct cpufreq_policy *policy;
343         unsigned int j;
344
345         if (!this_dbs_info->enable)
346                 return;
347
348         this_dbs_info->freq_lo = 0;
349         policy = this_dbs_info->cur_policy;
350
351         /*
352          * Every sampling_rate, we check, if current idle time is less
353          * than 20% (default), then we try to increase frequency
354          * Every sampling_rate, we look for a the lowest
355          * frequency which can sustain the load while keeping idle time over
356          * 30%. If such a frequency exist, we try to decrease to this frequency.
357          *
358          * Any frequency increase takes it to the maximum frequency.
359          * Frequency reduction happens at minimum steps of
360          * 5% (default) of current frequency
361          */
362
363         /* Get Absolute Load - in terms of freq */
364         max_load_freq = 0;
365
366         for_each_cpu_mask_nr(j, policy->cpus) {
367                 struct cpu_dbs_info_s *j_dbs_info;
368                 cputime64_t cur_wall_time, cur_idle_time;
369                 unsigned int idle_time, wall_time;
370                 unsigned int load, load_freq;
371                 int freq_avg;
372
373                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
374
375                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
376
377                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
378                                 j_dbs_info->prev_cpu_wall);
379                 j_dbs_info->prev_cpu_wall = cur_wall_time;
380
381                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
382                                 j_dbs_info->prev_cpu_idle);
383                 j_dbs_info->prev_cpu_idle = cur_idle_time;
384
385                 if (unlikely(!wall_time || wall_time < idle_time))
386                         continue;
387
388                 load = 100 * (wall_time - idle_time) / wall_time;
389
390                 freq_avg = __cpufreq_driver_getavg(policy, j);
391                 if (freq_avg <= 0)
392                         freq_avg = policy->cur;
393
394                 load_freq = load * freq_avg;
395                 if (load_freq > max_load_freq)
396                         max_load_freq = load_freq;
397         }
398
399         /* Check for frequency increase */
400         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
401                 /* if we are already at full speed then break out early */
402                 if (!dbs_tuners_ins.powersave_bias) {
403                         if (policy->cur == policy->max)
404                                 return;
405
406                         __cpufreq_driver_target(policy, policy->max,
407                                 CPUFREQ_RELATION_H);
408                 } else {
409                         int freq = powersave_bias_target(policy, policy->max,
410                                         CPUFREQ_RELATION_H);
411                         __cpufreq_driver_target(policy, freq,
412                                 CPUFREQ_RELATION_L);
413                 }
414                 return;
415         }
416
417         /* Check for frequency decrease */
418         /* if we cannot reduce the frequency anymore, break out early */
419         if (policy->cur == policy->min)
420                 return;
421
422         /*
423          * The optimal frequency is the frequency that is the lowest that
424          * can support the current CPU usage without triggering the up
425          * policy. To be safe, we focus 10 points under the threshold.
426          */
427         if (max_load_freq < (dbs_tuners_ins.up_threshold - 10) * policy->cur) {
428                 unsigned int freq_next;
429                 freq_next = max_load_freq / (dbs_tuners_ins.up_threshold - 10);
430
431                 if (!dbs_tuners_ins.powersave_bias) {
432                         __cpufreq_driver_target(policy, freq_next,
433                                         CPUFREQ_RELATION_L);
434                 } else {
435                         int freq = powersave_bias_target(policy, freq_next,
436                                         CPUFREQ_RELATION_L);
437                         __cpufreq_driver_target(policy, freq,
438                                 CPUFREQ_RELATION_L);
439                 }
440         }
441 }
442
443 static void do_dbs_timer(struct work_struct *work)
444 {
445         struct cpu_dbs_info_s *dbs_info =
446                 container_of(work, struct cpu_dbs_info_s, work.work);
447         unsigned int cpu = dbs_info->cpu;
448         int sample_type = dbs_info->sample_type;
449
450         /* We want all CPUs to do sampling nearly on same jiffy */
451         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
452
453         delay -= jiffies % delay;
454
455         if (lock_policy_rwsem_write(cpu) < 0)
456                 return;
457
458         if (!dbs_info->enable) {
459                 unlock_policy_rwsem_write(cpu);
460                 return;
461         }
462
463         /* Common NORMAL_SAMPLE setup */
464         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
465         if (!dbs_tuners_ins.powersave_bias ||
466             sample_type == DBS_NORMAL_SAMPLE) {
467                 dbs_check_cpu(dbs_info);
468                 if (dbs_info->freq_lo) {
469                         /* Setup timer for SUB_SAMPLE */
470                         dbs_info->sample_type = DBS_SUB_SAMPLE;
471                         delay = dbs_info->freq_hi_jiffies;
472                 }
473         } else {
474                 __cpufreq_driver_target(dbs_info->cur_policy,
475                                         dbs_info->freq_lo,
476                                         CPUFREQ_RELATION_H);
477         }
478         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
479         unlock_policy_rwsem_write(cpu);
480 }
481
482 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
483 {
484         /* We want all CPUs to do sampling nearly on same jiffy */
485         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
486         delay -= jiffies % delay;
487
488         dbs_info->enable = 1;
489         ondemand_powersave_bias_init();
490         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
491         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
492         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
493                               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(&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(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                 if (this_dbs_info->enable) /* Already enabled */
518                         break;
519
520                 mutex_lock(&dbs_mutex);
521                 dbs_enable++;
522
523                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
524                 if (rc) {
525                         dbs_enable--;
526                         mutex_unlock(&dbs_mutex);
527                         return rc;
528                 }
529
530                 for_each_cpu_mask_nr(j, policy->cpus) {
531                         struct cpu_dbs_info_s *j_dbs_info;
532                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
533                         j_dbs_info->cur_policy = policy;
534
535                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
536                                                 &j_dbs_info->prev_cpu_wall);
537                 }
538                 this_dbs_info->cpu = cpu;
539                 /*
540                  * Start the timerschedule work, when this governor
541                  * is used for first time
542                  */
543                 if (dbs_enable == 1) {
544                         unsigned int latency;
545                         /* policy latency is in nS. Convert it to uS first */
546                         latency = policy->cpuinfo.transition_latency / 1000;
547                         if (latency == 0)
548                                 latency = 1;
549
550                         def_sampling_rate = latency *
551                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
552
553                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
554                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
555
556                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
557                 }
558                 dbs_timer_init(this_dbs_info);
559
560                 mutex_unlock(&dbs_mutex);
561                 break;
562
563         case CPUFREQ_GOV_STOP:
564                 mutex_lock(&dbs_mutex);
565                 dbs_timer_exit(this_dbs_info);
566                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
567                 dbs_enable--;
568                 mutex_unlock(&dbs_mutex);
569
570                 break;
571
572         case CPUFREQ_GOV_LIMITS:
573                 mutex_lock(&dbs_mutex);
574                 if (policy->max < this_dbs_info->cur_policy->cur)
575                         __cpufreq_driver_target(this_dbs_info->cur_policy,
576                                                 policy->max,
577                                                 CPUFREQ_RELATION_H);
578                 else if (policy->min > this_dbs_info->cur_policy->cur)
579                         __cpufreq_driver_target(this_dbs_info->cur_policy,
580                                                 policy->min,
581                                                 CPUFREQ_RELATION_L);
582                 mutex_unlock(&dbs_mutex);
583                 break;
584         }
585         return 0;
586 }
587
588 struct cpufreq_governor cpufreq_gov_ondemand = {
589         .name                   = "ondemand",
590         .governor               = cpufreq_governor_dbs,
591         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
592         .owner                  = THIS_MODULE,
593 };
594 EXPORT_SYMBOL(cpufreq_gov_ondemand);
595
596 static int __init cpufreq_gov_dbs_init(void)
597 {
598         int err;
599
600         kondemand_wq = create_workqueue("kondemand");
601         if (!kondemand_wq) {
602                 printk(KERN_ERR "Creation of kondemand failed\n");
603                 return -EFAULT;
604         }
605         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
606         if (err)
607                 destroy_workqueue(kondemand_wq);
608
609         return err;
610 }
611
612 static void __exit cpufreq_gov_dbs_exit(void)
613 {
614         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
615         destroy_workqueue(kondemand_wq);
616 }
617
618
619 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
620 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
621 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
622                    "Low Latency Frequency Transition capable processors");
623 MODULE_LICENSE("GPL");
624
625 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
626 fs_initcall(cpufreq_gov_dbs_init);
627 #else
628 module_init(cpufreq_gov_dbs_init);
629 #endif
630 module_exit(cpufreq_gov_dbs_exit);