[CPUFREQ][2/6] cpufreq: Change load calculation in ondemand for software coordination
[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)
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 max_load_freq;
338
339         struct cpufreq_policy *policy;
340         unsigned int j;
341
342         if (!this_dbs_info->enable)
343                 return;
344
345         this_dbs_info->freq_lo = 0;
346         policy = this_dbs_info->cur_policy;
347
348         /*
349          * Every sampling_rate, we check, if current idle time is less
350          * than 20% (default), then we try to increase frequency
351          * Every sampling_rate, we look for a the lowest
352          * frequency which can sustain the load while keeping idle time over
353          * 30%. If such a frequency exist, we try to decrease to this frequency.
354          *
355          * Any frequency increase takes it to the maximum frequency.
356          * Frequency reduction happens at minimum steps of
357          * 5% (default) of current frequency
358          */
359
360         /* Get Absolute Load - in terms of freq */
361         max_load_freq = 0;
362
363         for_each_cpu_mask_nr(j, policy->cpus) {
364                 struct cpu_dbs_info_s *j_dbs_info;
365                 cputime64_t cur_wall_time, cur_idle_time;
366                 unsigned int idle_time, wall_time;
367                 unsigned int load, load_freq;
368                 int freq_avg;
369
370                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
371                 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
372                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
373                                 j_dbs_info->prev_cpu_wall);
374                 j_dbs_info->prev_cpu_wall = cur_wall_time;
375
376                 cur_idle_time = get_cpu_idle_time(j);
377                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
378                                 j_dbs_info->prev_cpu_idle);
379                 j_dbs_info->prev_cpu_idle = cur_idle_time;
380
381                 if (unlikely(wall_time <= idle_time ||
382                              (cputime_to_msecs(wall_time) <
383                               dbs_tuners_ins.sampling_rate / (2 * 1000)))) {
384                         continue;
385                 }
386
387                 load = 100 * (wall_time - idle_time) / wall_time;
388
389                 freq_avg = __cpufreq_driver_getavg(policy, j);
390                 if (freq_avg <= 0)
391                         freq_avg = policy->cur;
392
393                 load_freq = load * freq_avg;
394                 if (load_freq > max_load_freq)
395                         max_load_freq = load_freq;
396         }
397
398         /* Check for frequency increase */
399         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
400                 /* if we are already at full speed then break out early */
401                 if (!dbs_tuners_ins.powersave_bias) {
402                         if (policy->cur == policy->max)
403                                 return;
404
405                         __cpufreq_driver_target(policy, policy->max,
406                                 CPUFREQ_RELATION_H);
407                 } else {
408                         int freq = powersave_bias_target(policy, policy->max,
409                                         CPUFREQ_RELATION_H);
410                         __cpufreq_driver_target(policy, freq,
411                                 CPUFREQ_RELATION_L);
412                 }
413                 return;
414         }
415
416         /* Check for frequency decrease */
417         /* if we cannot reduce the frequency anymore, break out early */
418         if (policy->cur == policy->min)
419                 return;
420
421         /*
422          * The optimal frequency is the frequency that is the lowest that
423          * can support the current CPU usage without triggering the up
424          * policy. To be safe, we focus 10 points under the threshold.
425          */
426         if (max_load_freq < (dbs_tuners_ins.up_threshold - 10) * policy->cur) {
427                 unsigned int freq_next;
428                 freq_next = max_load_freq / (dbs_tuners_ins.up_threshold - 10);
429
430                 if (!dbs_tuners_ins.powersave_bias) {
431                         __cpufreq_driver_target(policy, freq_next,
432                                         CPUFREQ_RELATION_L);
433                 } else {
434                         int freq = powersave_bias_target(policy, freq_next,
435                                         CPUFREQ_RELATION_L);
436                         __cpufreq_driver_target(policy, freq,
437                                 CPUFREQ_RELATION_L);
438                 }
439         }
440 }
441
442 static void do_dbs_timer(struct work_struct *work)
443 {
444         struct cpu_dbs_info_s *dbs_info =
445                 container_of(work, struct cpu_dbs_info_s, work.work);
446         unsigned int cpu = dbs_info->cpu;
447         int sample_type = dbs_info->sample_type;
448
449         /* We want all CPUs to do sampling nearly on same jiffy */
450         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
451
452         delay -= jiffies % delay;
453
454         if (lock_policy_rwsem_write(cpu) < 0)
455                 return;
456
457         if (!dbs_info->enable) {
458                 unlock_policy_rwsem_write(cpu);
459                 return;
460         }
461
462         /* Common NORMAL_SAMPLE setup */
463         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
464         if (!dbs_tuners_ins.powersave_bias ||
465             sample_type == DBS_NORMAL_SAMPLE) {
466                 dbs_check_cpu(dbs_info);
467                 if (dbs_info->freq_lo) {
468                         /* Setup timer for SUB_SAMPLE */
469                         dbs_info->sample_type = DBS_SUB_SAMPLE;
470                         delay = dbs_info->freq_hi_jiffies;
471                 }
472         } else {
473                 __cpufreq_driver_target(dbs_info->cur_policy,
474                                         dbs_info->freq_lo,
475                                         CPUFREQ_RELATION_H);
476         }
477         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
478         unlock_policy_rwsem_write(cpu);
479 }
480
481 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
482 {
483         /* We want all CPUs to do sampling nearly on same jiffy */
484         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
485         delay -= jiffies % delay;
486
487         dbs_info->enable = 1;
488         ondemand_powersave_bias_init();
489         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
490         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
491         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
492                               delay);
493 }
494
495 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
496 {
497         dbs_info->enable = 0;
498         cancel_delayed_work(&dbs_info->work);
499 }
500
501 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
502                                    unsigned int event)
503 {
504         unsigned int cpu = policy->cpu;
505         struct cpu_dbs_info_s *this_dbs_info;
506         unsigned int j;
507         int rc;
508
509         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
510
511         switch (event) {
512         case CPUFREQ_GOV_START:
513                 if ((!cpu_online(cpu)) || (!policy->cur))
514                         return -EINVAL;
515
516                 if (this_dbs_info->enable) /* Already enabled */
517                         break;
518
519                 mutex_lock(&dbs_mutex);
520                 dbs_enable++;
521
522                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
523                 if (rc) {
524                         dbs_enable--;
525                         mutex_unlock(&dbs_mutex);
526                         return rc;
527                 }
528
529                 for_each_cpu_mask_nr(j, policy->cpus) {
530                         struct cpu_dbs_info_s *j_dbs_info;
531                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
532                         j_dbs_info->cur_policy = policy;
533
534                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
535                         j_dbs_info->prev_cpu_wall = get_jiffies_64();
536                 }
537                 this_dbs_info->cpu = cpu;
538                 /*
539                  * Start the timerschedule work, when this governor
540                  * is used for first time
541                  */
542                 if (dbs_enable == 1) {
543                         unsigned int latency;
544                         /* policy latency is in nS. Convert it to uS first */
545                         latency = policy->cpuinfo.transition_latency / 1000;
546                         if (latency == 0)
547                                 latency = 1;
548
549                         def_sampling_rate = latency *
550                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
551
552                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
553                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
554
555                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
556                 }
557                 dbs_timer_init(this_dbs_info);
558
559                 mutex_unlock(&dbs_mutex);
560                 break;
561
562         case CPUFREQ_GOV_STOP:
563                 mutex_lock(&dbs_mutex);
564                 dbs_timer_exit(this_dbs_info);
565                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
566                 dbs_enable--;
567                 mutex_unlock(&dbs_mutex);
568
569                 break;
570
571         case CPUFREQ_GOV_LIMITS:
572                 mutex_lock(&dbs_mutex);
573                 if (policy->max < this_dbs_info->cur_policy->cur)
574                         __cpufreq_driver_target(this_dbs_info->cur_policy,
575                                                 policy->max,
576                                                 CPUFREQ_RELATION_H);
577                 else if (policy->min > this_dbs_info->cur_policy->cur)
578                         __cpufreq_driver_target(this_dbs_info->cur_policy,
579                                                 policy->min,
580                                                 CPUFREQ_RELATION_L);
581                 mutex_unlock(&dbs_mutex);
582                 break;
583         }
584         return 0;
585 }
586
587 struct cpufreq_governor cpufreq_gov_ondemand = {
588         .name                   = "ondemand",
589         .governor               = cpufreq_governor_dbs,
590         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
591         .owner                  = THIS_MODULE,
592 };
593 EXPORT_SYMBOL(cpufreq_gov_ondemand);
594
595 static int __init cpufreq_gov_dbs_init(void)
596 {
597         int err;
598
599         kondemand_wq = create_workqueue("kondemand");
600         if (!kondemand_wq) {
601                 printk(KERN_ERR "Creation of kondemand failed\n");
602                 return -EFAULT;
603         }
604         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
605         if (err)
606                 destroy_workqueue(kondemand_wq);
607
608         return err;
609 }
610
611 static void __exit cpufreq_gov_dbs_exit(void)
612 {
613         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
614         destroy_workqueue(kondemand_wq);
615 }
616
617
618 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
619 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
620 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
621                    "Low Latency Frequency Transition capable processors");
622 MODULE_LICENSE("GPL");
623
624 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
625 fs_initcall(cpufreq_gov_dbs_init);
626 #else
627 module_init(cpufreq_gov_dbs_init);
628 #endif
629 module_exit(cpufreq_gov_dbs_exit);