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