mlx4_en: Memory leak on completion queue free
[linux-2.6.git] / drivers / cpufreq / cpufreq_conservative.c
1 /*
2  *  drivers/cpufreq/cpufreq_conservative.c
3  *
4  *  Copyright (C)  2001 Russell King
5  *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6  *                      Jun Nakajima <jun.nakajima@intel.com>
7  *            (C)  2004 Alexander Clouter <alex-kernel@digriz.org.uk>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License version 2 as
11  * published by the Free Software Foundation.
12  */
13
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/smp.h>
17 #include <linux/init.h>
18 #include <linux/interrupt.h>
19 #include <linux/ctype.h>
20 #include <linux/cpufreq.h>
21 #include <linux/sysctl.h>
22 #include <linux/types.h>
23 #include <linux/fs.h>
24 #include <linux/sysfs.h>
25 #include <linux/cpu.h>
26 #include <linux/kmod.h>
27 #include <linux/workqueue.h>
28 #include <linux/jiffies.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/percpu.h>
31 #include <linux/mutex.h>
32 /*
33  * dbs is used in this file as a shortform for demandbased switching
34  * It helps to keep variable names smaller, simpler
35  */
36
37 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
38 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
39
40 /*
41  * The polling frequency of this governor depends on the capability of
42  * the processor. Default polling frequency is 1000 times the transition
43  * latency of the processor. The governor will work on any processor with
44  * transition latency <= 10mS, using appropriate sampling
45  * rate.
46  * For CPUs with transition latency > 10mS (mostly drivers
47  * with CPUFREQ_ETERNAL), this governor will not work.
48  * All times here are in uS.
49  */
50 static unsigned int def_sampling_rate;
51 #define MIN_SAMPLING_RATE_RATIO                 (2)
52 /* for correct statistics, we need at least 10 ticks between each measure */
53 #define MIN_STAT_SAMPLING_RATE                  \
54                         (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
55 #define MIN_SAMPLING_RATE                       \
56                         (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
57 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
58 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
59 #define DEF_SAMPLING_DOWN_FACTOR                (1)
60 #define MAX_SAMPLING_DOWN_FACTOR                (10)
61 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
62
63 static void do_dbs_timer(struct work_struct *work);
64
65 struct cpu_dbs_info_s {
66         struct cpufreq_policy *cur_policy;
67         unsigned int prev_cpu_idle_up;
68         unsigned int prev_cpu_idle_down;
69         unsigned int enable;
70         unsigned int down_skip;
71         unsigned int requested_freq;
72 };
73 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
74
75 static unsigned int dbs_enable; /* number of CPUs using this policy */
76
77 /*
78  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
79  * lock and dbs_mutex. cpu_hotplug lock should always be held before
80  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
81  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
82  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
83  * is recursive for the same process. -Venki
84  */
85 static DEFINE_MUTEX (dbs_mutex);
86 static DECLARE_DELAYED_WORK(dbs_work, do_dbs_timer);
87
88 struct dbs_tuners {
89         unsigned int sampling_rate;
90         unsigned int sampling_down_factor;
91         unsigned int up_threshold;
92         unsigned int down_threshold;
93         unsigned int ignore_nice;
94         unsigned int freq_step;
95 };
96
97 static struct dbs_tuners dbs_tuners_ins = {
98         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
99         .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
100         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
101         .ignore_nice = 0,
102         .freq_step = 5,
103 };
104
105 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
106 {
107         unsigned int add_nice = 0, ret;
108
109         if (dbs_tuners_ins.ignore_nice)
110                 add_nice = kstat_cpu(cpu).cpustat.nice;
111
112         ret = kstat_cpu(cpu).cpustat.idle +
113                 kstat_cpu(cpu).cpustat.iowait +
114                 add_nice;
115
116         return ret;
117 }
118
119 /* keep track of frequency transitions */
120 static int
121 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
122                      void *data)
123 {
124         struct cpufreq_freqs *freq = data;
125         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info,
126                                                         freq->cpu);
127
128         if (!this_dbs_info->enable)
129                 return 0;
130
131         this_dbs_info->requested_freq = freq->new;
132
133         return 0;
134 }
135
136 static struct notifier_block dbs_cpufreq_notifier_block = {
137         .notifier_call = dbs_cpufreq_notifier
138 };
139
140 /************************** sysfs interface ************************/
141 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
142 {
143         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
144 }
145
146 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
147 {
148         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
149 }
150
151 #define define_one_ro(_name)                            \
152 static struct freq_attr _name =                         \
153 __ATTR(_name, 0444, show_##_name, NULL)
154
155 define_one_ro(sampling_rate_max);
156 define_one_ro(sampling_rate_min);
157
158 /* cpufreq_conservative Governor Tunables */
159 #define show_one(file_name, object)                                     \
160 static ssize_t show_##file_name                                         \
161 (struct cpufreq_policy *unused, char *buf)                              \
162 {                                                                       \
163         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
164 }
165 show_one(sampling_rate, sampling_rate);
166 show_one(sampling_down_factor, sampling_down_factor);
167 show_one(up_threshold, up_threshold);
168 show_one(down_threshold, down_threshold);
169 show_one(ignore_nice_load, ignore_nice);
170 show_one(freq_step, freq_step);
171
172 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
173                 const char *buf, size_t count)
174 {
175         unsigned int input;
176         int ret;
177         ret = sscanf (buf, "%u", &input);
178         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
179                 return -EINVAL;
180
181         mutex_lock(&dbs_mutex);
182         dbs_tuners_ins.sampling_down_factor = input;
183         mutex_unlock(&dbs_mutex);
184
185         return count;
186 }
187
188 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
189                 const char *buf, size_t count)
190 {
191         unsigned int input;
192         int ret;
193         ret = sscanf (buf, "%u", &input);
194
195         mutex_lock(&dbs_mutex);
196         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
197                 mutex_unlock(&dbs_mutex);
198                 return -EINVAL;
199         }
200
201         dbs_tuners_ins.sampling_rate = input;
202         mutex_unlock(&dbs_mutex);
203
204         return count;
205 }
206
207 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
208                 const char *buf, size_t count)
209 {
210         unsigned int input;
211         int ret;
212         ret = sscanf (buf, "%u", &input);
213
214         mutex_lock(&dbs_mutex);
215         if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) {
216                 mutex_unlock(&dbs_mutex);
217                 return -EINVAL;
218         }
219
220         dbs_tuners_ins.up_threshold = input;
221         mutex_unlock(&dbs_mutex);
222
223         return count;
224 }
225
226 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
227                 const char *buf, size_t count)
228 {
229         unsigned int input;
230         int ret;
231         ret = sscanf (buf, "%u", &input);
232
233         mutex_lock(&dbs_mutex);
234         if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
235                 mutex_unlock(&dbs_mutex);
236                 return -EINVAL;
237         }
238
239         dbs_tuners_ins.down_threshold = input;
240         mutex_unlock(&dbs_mutex);
241
242         return count;
243 }
244
245 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
246                 const char *buf, size_t count)
247 {
248         unsigned int input;
249         int ret;
250
251         unsigned int j;
252
253         ret = sscanf(buf, "%u", &input);
254         if (ret != 1)
255                 return -EINVAL;
256
257         if (input > 1)
258                 input = 1;
259
260         mutex_lock(&dbs_mutex);
261         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
262                 mutex_unlock(&dbs_mutex);
263                 return count;
264         }
265         dbs_tuners_ins.ignore_nice = input;
266
267         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
268         for_each_online_cpu(j) {
269                 struct cpu_dbs_info_s *j_dbs_info;
270                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
271                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
272                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
273         }
274         mutex_unlock(&dbs_mutex);
275
276         return count;
277 }
278
279 static ssize_t store_freq_step(struct cpufreq_policy *policy,
280                 const char *buf, size_t count)
281 {
282         unsigned int input;
283         int ret;
284
285         ret = sscanf(buf, "%u", &input);
286
287         if (ret != 1)
288                 return -EINVAL;
289
290         if (input > 100)
291                 input = 100;
292
293         /* no need to test here if freq_step is zero as the user might actually
294          * want this, they would be crazy though :) */
295         mutex_lock(&dbs_mutex);
296         dbs_tuners_ins.freq_step = input;
297         mutex_unlock(&dbs_mutex);
298
299         return count;
300 }
301
302 #define define_one_rw(_name) \
303 static struct freq_attr _name = \
304 __ATTR(_name, 0644, show_##_name, store_##_name)
305
306 define_one_rw(sampling_rate);
307 define_one_rw(sampling_down_factor);
308 define_one_rw(up_threshold);
309 define_one_rw(down_threshold);
310 define_one_rw(ignore_nice_load);
311 define_one_rw(freq_step);
312
313 static struct attribute * dbs_attributes[] = {
314         &sampling_rate_max.attr,
315         &sampling_rate_min.attr,
316         &sampling_rate.attr,
317         &sampling_down_factor.attr,
318         &up_threshold.attr,
319         &down_threshold.attr,
320         &ignore_nice_load.attr,
321         &freq_step.attr,
322         NULL
323 };
324
325 static struct attribute_group dbs_attr_group = {
326         .attrs = dbs_attributes,
327         .name = "conservative",
328 };
329
330 /************************** sysfs end ************************/
331
332 static void dbs_check_cpu(int cpu)
333 {
334         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
335         unsigned int tmp_idle_ticks, total_idle_ticks;
336         unsigned int freq_target;
337         unsigned int freq_down_sampling_rate;
338         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
339         struct cpufreq_policy *policy;
340
341         if (!this_dbs_info->enable)
342                 return;
343
344         policy = this_dbs_info->cur_policy;
345
346         /*
347          * The default safe range is 20% to 80%
348          * Every sampling_rate, we check
349          *      - If current idle time is less than 20%, then we try to
350          *        increase frequency
351          * Every sampling_rate*sampling_down_factor, we check
352          *      - If current idle time is more than 80%, then we try to
353          *        decrease frequency
354          *
355          * Any frequency increase takes it to the maximum frequency.
356          * Frequency reduction happens at minimum steps of
357          * 5% (default) of max_frequency
358          */
359
360         /* Check for frequency increase */
361         idle_ticks = UINT_MAX;
362
363         /* Check for frequency increase */
364         total_idle_ticks = get_cpu_idle_time(cpu);
365         tmp_idle_ticks = total_idle_ticks -
366                 this_dbs_info->prev_cpu_idle_up;
367         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
368
369         if (tmp_idle_ticks < idle_ticks)
370                 idle_ticks = tmp_idle_ticks;
371
372         /* Scale idle ticks by 100 and compare with up and down ticks */
373         idle_ticks *= 100;
374         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
375                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
376
377         if (idle_ticks < up_idle_ticks) {
378                 this_dbs_info->down_skip = 0;
379                 this_dbs_info->prev_cpu_idle_down =
380                         this_dbs_info->prev_cpu_idle_up;
381
382                 /* if we are already at full speed then break out early */
383                 if (this_dbs_info->requested_freq == policy->max)
384                         return;
385
386                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
387
388                 /* max freq cannot be less than 100. But who knows.... */
389                 if (unlikely(freq_target == 0))
390                         freq_target = 5;
391
392                 this_dbs_info->requested_freq += freq_target;
393                 if (this_dbs_info->requested_freq > policy->max)
394                         this_dbs_info->requested_freq = policy->max;
395
396                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
397                         CPUFREQ_RELATION_H);
398                 return;
399         }
400
401         /* Check for frequency decrease */
402         this_dbs_info->down_skip++;
403         if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
404                 return;
405
406         /* Check for frequency decrease */
407         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
408         tmp_idle_ticks = total_idle_ticks -
409                 this_dbs_info->prev_cpu_idle_down;
410         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
411
412         if (tmp_idle_ticks < idle_ticks)
413                 idle_ticks = tmp_idle_ticks;
414
415         /* Scale idle ticks by 100 and compare with up and down ticks */
416         idle_ticks *= 100;
417         this_dbs_info->down_skip = 0;
418
419         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
420                 dbs_tuners_ins.sampling_down_factor;
421         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
422                 usecs_to_jiffies(freq_down_sampling_rate);
423
424         if (idle_ticks > down_idle_ticks) {
425                 /*
426                  * if we are already at the lowest speed then break out early
427                  * or if we 'cannot' reduce the speed as the user might want
428                  * freq_target to be zero
429                  */
430                 if (this_dbs_info->requested_freq == policy->min
431                                 || dbs_tuners_ins.freq_step == 0)
432                         return;
433
434                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
435
436                 /* max freq cannot be less than 100. But who knows.... */
437                 if (unlikely(freq_target == 0))
438                         freq_target = 5;
439
440                 this_dbs_info->requested_freq -= freq_target;
441                 if (this_dbs_info->requested_freq < policy->min)
442                         this_dbs_info->requested_freq = policy->min;
443
444                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
445                                 CPUFREQ_RELATION_H);
446                 return;
447         }
448 }
449
450 static void do_dbs_timer(struct work_struct *work)
451 {
452         int i;
453         mutex_lock(&dbs_mutex);
454         for_each_online_cpu(i)
455                 dbs_check_cpu(i);
456         schedule_delayed_work(&dbs_work,
457                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
458         mutex_unlock(&dbs_mutex);
459 }
460
461 static inline void dbs_timer_init(void)
462 {
463         init_timer_deferrable(&dbs_work.timer);
464         schedule_delayed_work(&dbs_work,
465                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
466         return;
467 }
468
469 static inline void dbs_timer_exit(void)
470 {
471         cancel_delayed_work(&dbs_work);
472         return;
473 }
474
475 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
476                                    unsigned int event)
477 {
478         unsigned int cpu = policy->cpu;
479         struct cpu_dbs_info_s *this_dbs_info;
480         unsigned int j;
481         int rc;
482
483         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
484
485         switch (event) {
486         case CPUFREQ_GOV_START:
487                 if ((!cpu_online(cpu)) || (!policy->cur))
488                         return -EINVAL;
489
490                 if (this_dbs_info->enable) /* Already enabled */
491                         break;
492
493                 mutex_lock(&dbs_mutex);
494
495                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
496                 if (rc) {
497                         mutex_unlock(&dbs_mutex);
498                         return rc;
499                 }
500
501                 for_each_cpu_mask_nr(j, policy->cpus) {
502                         struct cpu_dbs_info_s *j_dbs_info;
503                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
504                         j_dbs_info->cur_policy = policy;
505
506                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
507                         j_dbs_info->prev_cpu_idle_down
508                                 = j_dbs_info->prev_cpu_idle_up;
509                 }
510                 this_dbs_info->enable = 1;
511                 this_dbs_info->down_skip = 0;
512                 this_dbs_info->requested_freq = policy->cur;
513
514                 dbs_enable++;
515                 /*
516                  * Start the timerschedule work, when this governor
517                  * is used for first time
518                  */
519                 if (dbs_enable == 1) {
520                         unsigned int latency;
521                         /* policy latency is in nS. Convert it to uS first */
522                         latency = policy->cpuinfo.transition_latency / 1000;
523                         if (latency == 0)
524                                 latency = 1;
525
526                         def_sampling_rate = 10 * latency *
527                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
528
529                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
530                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
531
532                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
533
534                         dbs_timer_init();
535                         cpufreq_register_notifier(
536                                         &dbs_cpufreq_notifier_block,
537                                         CPUFREQ_TRANSITION_NOTIFIER);
538                 }
539
540                 mutex_unlock(&dbs_mutex);
541                 break;
542
543         case CPUFREQ_GOV_STOP:
544                 mutex_lock(&dbs_mutex);
545                 this_dbs_info->enable = 0;
546                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
547                 dbs_enable--;
548                 /*
549                  * Stop the timerschedule work, when this governor
550                  * is used for first time
551                  */
552                 if (dbs_enable == 0) {
553                         dbs_timer_exit();
554                         cpufreq_unregister_notifier(
555                                         &dbs_cpufreq_notifier_block,
556                                         CPUFREQ_TRANSITION_NOTIFIER);
557                 }
558
559                 mutex_unlock(&dbs_mutex);
560
561                 break;
562
563         case CPUFREQ_GOV_LIMITS:
564                 mutex_lock(&dbs_mutex);
565                 if (policy->max < this_dbs_info->cur_policy->cur)
566                         __cpufreq_driver_target(
567                                         this_dbs_info->cur_policy,
568                                         policy->max, CPUFREQ_RELATION_H);
569                 else if (policy->min > this_dbs_info->cur_policy->cur)
570                         __cpufreq_driver_target(
571                                         this_dbs_info->cur_policy,
572                                         policy->min, CPUFREQ_RELATION_L);
573                 mutex_unlock(&dbs_mutex);
574                 break;
575         }
576         return 0;
577 }
578
579 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
580 static
581 #endif
582 struct cpufreq_governor cpufreq_gov_conservative = {
583         .name                   = "conservative",
584         .governor               = cpufreq_governor_dbs,
585         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
586         .owner                  = THIS_MODULE,
587 };
588
589 static int __init cpufreq_gov_dbs_init(void)
590 {
591         return cpufreq_register_governor(&cpufreq_gov_conservative);
592 }
593
594 static void __exit cpufreq_gov_dbs_exit(void)
595 {
596         /* Make sure that the scheduled work is indeed not running */
597         flush_scheduled_work();
598
599         cpufreq_unregister_governor(&cpufreq_gov_conservative);
600 }
601
602
603 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
604 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
605                 "Low Latency Frequency Transition capable processors "
606                 "optimised for use in a battery environment");
607 MODULE_LICENSE ("GPL");
608
609 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
610 fs_initcall(cpufreq_gov_dbs_init);
611 #else
612 module_init(cpufreq_gov_dbs_init);
613 #endif
614 module_exit(cpufreq_gov_dbs_exit);