Automatic merge of rsync://rsync.kernel.org/pub/scm/linux/kernel/git/gregkh/pci-2.6
[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/smp.h>
16 #include <linux/init.h>
17 #include <linux/interrupt.h>
18 #include <linux/ctype.h>
19 #include <linux/cpufreq.h>
20 #include <linux/sysctl.h>
21 #include <linux/types.h>
22 #include <linux/fs.h>
23 #include <linux/sysfs.h>
24 #include <linux/sched.h>
25 #include <linux/kmod.h>
26 #include <linux/workqueue.h>
27 #include <linux/jiffies.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/percpu.h>
30
31 /*
32  * dbs is used in this file as a shortform for demandbased switching
33  * It helps to keep variable names smaller, simpler
34  */
35
36 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
37 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
38 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
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 with CPUFREQ_ETERNAL)
47  * 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                       (def_sampling_rate / 2)
52 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
53 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
54 #define DEF_SAMPLING_DOWN_FACTOR                (1)
55 #define MAX_SAMPLING_DOWN_FACTOR                (10)
56 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
57
58 static void do_dbs_timer(void *data);
59
60 struct cpu_dbs_info_s {
61         struct cpufreq_policy   *cur_policy;
62         unsigned int            prev_cpu_idle_up;
63         unsigned int            prev_cpu_idle_down;
64         unsigned int            enable;
65 };
66 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
67
68 static unsigned int dbs_enable; /* number of CPUs using this policy */
69
70 static DECLARE_MUTEX    (dbs_sem);
71 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
72
73 struct dbs_tuners {
74         unsigned int            sampling_rate;
75         unsigned int            sampling_down_factor;
76         unsigned int            up_threshold;
77         unsigned int            ignore_nice;
78 };
79
80 static struct dbs_tuners dbs_tuners_ins = {
81         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
82         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
83 };
84
85 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
86 {
87         return  kstat_cpu(cpu).cpustat.idle +
88                 kstat_cpu(cpu).cpustat.iowait +
89                 ( !dbs_tuners_ins.ignore_nice ? 
90                   kstat_cpu(cpu).cpustat.nice :
91                   0);
92 }
93
94 /************************** sysfs interface ************************/
95 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
96 {
97         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
98 }
99
100 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
101 {
102         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
103 }
104
105 #define define_one_ro(_name)                                    \
106 static struct freq_attr _name =                                 \
107 __ATTR(_name, 0444, show_##_name, NULL)
108
109 define_one_ro(sampling_rate_max);
110 define_one_ro(sampling_rate_min);
111
112 /* cpufreq_ondemand Governor Tunables */
113 #define show_one(file_name, object)                                     \
114 static ssize_t show_##file_name                                         \
115 (struct cpufreq_policy *unused, char *buf)                              \
116 {                                                                       \
117         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
118 }
119 show_one(sampling_rate, sampling_rate);
120 show_one(sampling_down_factor, sampling_down_factor);
121 show_one(up_threshold, up_threshold);
122 show_one(ignore_nice, ignore_nice);
123
124 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
125                 const char *buf, size_t count)
126 {
127         unsigned int input;
128         int ret;
129         ret = sscanf (buf, "%u", &input);
130         if (ret != 1 )
131                 return -EINVAL;
132
133         if (input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
134                 return -EINVAL;
135
136         down(&dbs_sem);
137         dbs_tuners_ins.sampling_down_factor = input;
138         up(&dbs_sem);
139
140         return count;
141 }
142
143 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
144                 const char *buf, size_t count)
145 {
146         unsigned int input;
147         int ret;
148         ret = sscanf (buf, "%u", &input);
149
150         down(&dbs_sem);
151         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
152                 up(&dbs_sem);
153                 return -EINVAL;
154         }
155
156         dbs_tuners_ins.sampling_rate = input;
157         up(&dbs_sem);
158
159         return count;
160 }
161
162 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
163                 const char *buf, size_t count)
164 {
165         unsigned int input;
166         int ret;
167         ret = sscanf (buf, "%u", &input);
168
169         down(&dbs_sem);
170         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 
171                         input < MIN_FREQUENCY_UP_THRESHOLD) {
172                 up(&dbs_sem);
173                 return -EINVAL;
174         }
175
176         dbs_tuners_ins.up_threshold = input;
177         up(&dbs_sem);
178
179         return count;
180 }
181
182 static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
183                 const char *buf, size_t count)
184 {
185         unsigned int input;
186         int ret;
187
188         unsigned int j;
189         
190         ret = sscanf (buf, "%u", &input);
191         if ( ret != 1 )
192                 return -EINVAL;
193
194         if ( input > 1 )
195                 input = 1;
196         
197         down(&dbs_sem);
198         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
199                 up(&dbs_sem);
200                 return count;
201         }
202         dbs_tuners_ins.ignore_nice = input;
203
204         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
205         for_each_online_cpu(j) {
206                 struct cpu_dbs_info_s *j_dbs_info;
207                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
208                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
209                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
210         }
211         up(&dbs_sem);
212
213         return count;
214 }
215
216 #define define_one_rw(_name) \
217 static struct freq_attr _name = \
218 __ATTR(_name, 0644, show_##_name, store_##_name)
219
220 define_one_rw(sampling_rate);
221 define_one_rw(sampling_down_factor);
222 define_one_rw(up_threshold);
223 define_one_rw(ignore_nice);
224
225 static struct attribute * dbs_attributes[] = {
226         &sampling_rate_max.attr,
227         &sampling_rate_min.attr,
228         &sampling_rate.attr,
229         &sampling_down_factor.attr,
230         &up_threshold.attr,
231         &ignore_nice.attr,
232         NULL
233 };
234
235 static struct attribute_group dbs_attr_group = {
236         .attrs = dbs_attributes,
237         .name = "ondemand",
238 };
239
240 /************************** sysfs end ************************/
241
242 static void dbs_check_cpu(int cpu)
243 {
244         unsigned int idle_ticks, up_idle_ticks, total_ticks;
245         unsigned int freq_next;
246         unsigned int freq_down_sampling_rate;
247         static int down_skip[NR_CPUS];
248         struct cpu_dbs_info_s *this_dbs_info;
249
250         struct cpufreq_policy *policy;
251         unsigned int j;
252
253         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
254         if (!this_dbs_info->enable)
255                 return;
256
257         policy = this_dbs_info->cur_policy;
258         /* 
259          * Every sampling_rate, we check, if current idle time is less
260          * than 20% (default), then we try to increase frequency
261          * Every sampling_rate*sampling_down_factor, we look for a the lowest
262          * frequency which can sustain the load while keeping idle time over
263          * 30%. If such a frequency exist, we try to decrease to this frequency.
264          *
265          * Any frequency increase takes it to the maximum frequency. 
266          * Frequency reduction happens at minimum steps of 
267          * 5% (default) of current frequency 
268          */
269
270         /* Check for frequency increase */
271         idle_ticks = UINT_MAX;
272         for_each_cpu_mask(j, policy->cpus) {
273                 unsigned int tmp_idle_ticks, total_idle_ticks;
274                 struct cpu_dbs_info_s *j_dbs_info;
275
276                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
277                 total_idle_ticks = get_cpu_idle_time(j);
278                 tmp_idle_ticks = total_idle_ticks -
279                         j_dbs_info->prev_cpu_idle_up;
280                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
281
282                 if (tmp_idle_ticks < idle_ticks)
283                         idle_ticks = tmp_idle_ticks;
284         }
285
286         /* Scale idle ticks by 100 and compare with up and down ticks */
287         idle_ticks *= 100;
288         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
289                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
290
291         if (idle_ticks < up_idle_ticks) {
292                 down_skip[cpu] = 0;
293                 for_each_cpu_mask(j, policy->cpus) {
294                         struct cpu_dbs_info_s *j_dbs_info;
295
296                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
297                         j_dbs_info->prev_cpu_idle_down = 
298                                         j_dbs_info->prev_cpu_idle_up;
299                 }
300                 /* if we are already at full speed then break out early */
301                 if (policy->cur == policy->max)
302                         return;
303                 
304                 __cpufreq_driver_target(policy, policy->max, 
305                         CPUFREQ_RELATION_H);
306                 return;
307         }
308
309         /* Check for frequency decrease */
310         down_skip[cpu]++;
311         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
312                 return;
313
314         idle_ticks = UINT_MAX;
315         for_each_cpu_mask(j, policy->cpus) {
316                 unsigned int tmp_idle_ticks, total_idle_ticks;
317                 struct cpu_dbs_info_s *j_dbs_info;
318
319                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
320                 /* Check for frequency decrease */
321                 total_idle_ticks = j_dbs_info->prev_cpu_idle_up;
322                 tmp_idle_ticks = total_idle_ticks -
323                         j_dbs_info->prev_cpu_idle_down;
324                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
325
326                 if (tmp_idle_ticks < idle_ticks)
327                         idle_ticks = tmp_idle_ticks;
328         }
329
330         down_skip[cpu] = 0;
331         /* if we cannot reduce the frequency anymore, break out early */
332         if (policy->cur == policy->min)
333                 return;
334
335         /* Compute how many ticks there are between two measurements */
336         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
337                 dbs_tuners_ins.sampling_down_factor;
338         total_ticks = usecs_to_jiffies(freq_down_sampling_rate);
339
340         /*
341          * The optimal frequency is the frequency that is the lowest that
342          * can support the current CPU usage without triggering the up
343          * policy. To be safe, we focus 10 points under the threshold.
344          */
345         freq_next = ((total_ticks - idle_ticks) * 100) / total_ticks;
346         freq_next = (freq_next * policy->cur) / 
347                         (dbs_tuners_ins.up_threshold - 10);
348
349         if (freq_next <= ((policy->cur * 95) / 100))
350                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
351 }
352
353 static void do_dbs_timer(void *data)
354
355         int i;
356         down(&dbs_sem);
357         for_each_online_cpu(i)
358                 dbs_check_cpu(i);
359         schedule_delayed_work(&dbs_work, 
360                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
361         up(&dbs_sem);
362
363
364 static inline void dbs_timer_init(void)
365 {
366         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
367         schedule_delayed_work(&dbs_work,
368                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
369         return;
370 }
371
372 static inline void dbs_timer_exit(void)
373 {
374         cancel_delayed_work(&dbs_work);
375         return;
376 }
377
378 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
379                                    unsigned int event)
380 {
381         unsigned int cpu = policy->cpu;
382         struct cpu_dbs_info_s *this_dbs_info;
383         unsigned int j;
384
385         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
386
387         switch (event) {
388         case CPUFREQ_GOV_START:
389                 if ((!cpu_online(cpu)) || 
390                     (!policy->cur))
391                         return -EINVAL;
392
393                 if (policy->cpuinfo.transition_latency >
394                                 (TRANSITION_LATENCY_LIMIT * 1000))
395                         return -EINVAL;
396                 if (this_dbs_info->enable) /* Already enabled */
397                         break;
398                  
399                 down(&dbs_sem);
400                 for_each_cpu_mask(j, policy->cpus) {
401                         struct cpu_dbs_info_s *j_dbs_info;
402                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
403                         j_dbs_info->cur_policy = policy;
404                 
405                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
406                         j_dbs_info->prev_cpu_idle_down
407                                 = j_dbs_info->prev_cpu_idle_up;
408                 }
409                 this_dbs_info->enable = 1;
410                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
411                 dbs_enable++;
412                 /*
413                  * Start the timerschedule work, when this governor
414                  * is used for first time
415                  */
416                 if (dbs_enable == 1) {
417                         unsigned int latency;
418                         /* policy latency is in nS. Convert it to uS first */
419
420                         latency = policy->cpuinfo.transition_latency;
421                         if (latency < 1000)
422                                 latency = 1000;
423
424                         def_sampling_rate = (latency / 1000) *
425                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
426                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
427                         dbs_tuners_ins.ignore_nice = 0;
428
429                         dbs_timer_init();
430                 }
431                 
432                 up(&dbs_sem);
433                 break;
434
435         case CPUFREQ_GOV_STOP:
436                 down(&dbs_sem);
437                 this_dbs_info->enable = 0;
438                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
439                 dbs_enable--;
440                 /*
441                  * Stop the timerschedule work, when this governor
442                  * is used for first time
443                  */
444                 if (dbs_enable == 0) 
445                         dbs_timer_exit();
446                 
447                 up(&dbs_sem);
448
449                 break;
450
451         case CPUFREQ_GOV_LIMITS:
452                 down(&dbs_sem);
453                 if (policy->max < this_dbs_info->cur_policy->cur)
454                         __cpufreq_driver_target(
455                                         this_dbs_info->cur_policy,
456                                         policy->max, CPUFREQ_RELATION_H);
457                 else if (policy->min > this_dbs_info->cur_policy->cur)
458                         __cpufreq_driver_target(
459                                         this_dbs_info->cur_policy,
460                                         policy->min, CPUFREQ_RELATION_L);
461                 up(&dbs_sem);
462                 break;
463         }
464         return 0;
465 }
466
467 static struct cpufreq_governor cpufreq_gov_dbs = {
468         .name           = "ondemand",
469         .governor       = cpufreq_governor_dbs,
470         .owner          = THIS_MODULE,
471 };
472
473 static int __init cpufreq_gov_dbs_init(void)
474 {
475         return cpufreq_register_governor(&cpufreq_gov_dbs);
476 }
477
478 static void __exit cpufreq_gov_dbs_exit(void)
479 {
480         /* Make sure that the scheduled work is indeed not running */
481         flush_scheduled_work();
482
483         cpufreq_unregister_governor(&cpufreq_gov_dbs);
484 }
485
486
487 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
488 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
489                 "Low Latency Frequency Transition capable processors");
490 MODULE_LICENSE ("GPL");
491
492 module_init(cpufreq_gov_dbs_init);
493 module_exit(cpufreq_gov_dbs_exit);