Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu
[linux-3.10.git] / drivers / cpufreq / cpufreq.c
1 /*
2  *  linux/drivers/cpufreq/cpufreq.c
3  *
4  *  Copyright (C) 2001 Russell King
5  *            (C) 2002 - 2003 Dominik Brodowski <linux@brodo.de>
6  *
7  *  Oct 2005 - Ashok Raj <ashok.raj@intel.com>
8  *      Added handling for CPU hotplug
9  *  Feb 2006 - Jacob Shin <jacob.shin@amd.com>
10  *      Fix handling for CPU hotplug -- affected CPUs
11  *
12  * This program is free software; you can redistribute it and/or modify
13  * it under the terms of the GNU General Public License version 2 as
14  * published by the Free Software Foundation.
15  *
16  */
17
18 #include <linux/kernel.h>
19 #include <linux/module.h>
20 #include <linux/init.h>
21 #include <linux/notifier.h>
22 #include <linux/cpufreq.h>
23 #include <linux/delay.h>
24 #include <linux/interrupt.h>
25 #include <linux/spinlock.h>
26 #include <linux/device.h>
27 #include <linux/slab.h>
28 #include <linux/cpu.h>
29 #include <linux/completion.h>
30 #include <linux/mutex.h>
31
32 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_CORE, \
33                                                 "cpufreq-core", msg)
34
35 /**
36  * The "cpufreq driver" - the arch- or hardware-dependent low
37  * level driver of CPUFreq support, and its spinlock. This lock
38  * also protects the cpufreq_cpu_data array.
39  */
40 static struct cpufreq_driver *cpufreq_driver;
41 static DEFINE_PER_CPU(struct cpufreq_policy *, cpufreq_cpu_data);
42 #ifdef CONFIG_HOTPLUG_CPU
43 /* This one keeps track of the previously set governor of a removed CPU */
44 static DEFINE_PER_CPU(char[CPUFREQ_NAME_LEN], cpufreq_cpu_governor);
45 #endif
46 static DEFINE_SPINLOCK(cpufreq_driver_lock);
47
48 /*
49  * cpu_policy_rwsem is a per CPU reader-writer semaphore designed to cure
50  * all cpufreq/hotplug/workqueue/etc related lock issues.
51  *
52  * The rules for this semaphore:
53  * - Any routine that wants to read from the policy structure will
54  *   do a down_read on this semaphore.
55  * - Any routine that will write to the policy structure and/or may take away
56  *   the policy altogether (eg. CPU hotplug), will hold this lock in write
57  *   mode before doing so.
58  *
59  * Additional rules:
60  * - All holders of the lock should check to make sure that the CPU they
61  *   are concerned with are online after they get the lock.
62  * - Governor routines that can be called in cpufreq hotplug path should not
63  *   take this sem as top level hotplug notifier handler takes this.
64  * - Lock should not be held across
65  *     __cpufreq_governor(data, CPUFREQ_GOV_STOP);
66  */
67 static DEFINE_PER_CPU(int, cpufreq_policy_cpu);
68 static DEFINE_PER_CPU(struct rw_semaphore, cpu_policy_rwsem);
69
70 #define lock_policy_rwsem(mode, cpu)                                    \
71 int lock_policy_rwsem_##mode                                            \
72 (int cpu)                                                               \
73 {                                                                       \
74         int policy_cpu = per_cpu(cpufreq_policy_cpu, cpu);              \
75         BUG_ON(policy_cpu == -1);                                       \
76         down_##mode(&per_cpu(cpu_policy_rwsem, policy_cpu));            \
77         if (unlikely(!cpu_online(cpu))) {                               \
78                 up_##mode(&per_cpu(cpu_policy_rwsem, policy_cpu));      \
79                 return -1;                                              \
80         }                                                               \
81                                                                         \
82         return 0;                                                       \
83 }
84
85 lock_policy_rwsem(read, cpu);
86 EXPORT_SYMBOL_GPL(lock_policy_rwsem_read);
87
88 lock_policy_rwsem(write, cpu);
89 EXPORT_SYMBOL_GPL(lock_policy_rwsem_write);
90
91 void unlock_policy_rwsem_read(int cpu)
92 {
93         int policy_cpu = per_cpu(cpufreq_policy_cpu, cpu);
94         BUG_ON(policy_cpu == -1);
95         up_read(&per_cpu(cpu_policy_rwsem, policy_cpu));
96 }
97 EXPORT_SYMBOL_GPL(unlock_policy_rwsem_read);
98
99 void unlock_policy_rwsem_write(int cpu)
100 {
101         int policy_cpu = per_cpu(cpufreq_policy_cpu, cpu);
102         BUG_ON(policy_cpu == -1);
103         up_write(&per_cpu(cpu_policy_rwsem, policy_cpu));
104 }
105 EXPORT_SYMBOL_GPL(unlock_policy_rwsem_write);
106
107
108 /* internal prototypes */
109 static int __cpufreq_governor(struct cpufreq_policy *policy,
110                 unsigned int event);
111 static unsigned int __cpufreq_get(unsigned int cpu);
112 static void handle_update(struct work_struct *work);
113
114 /**
115  * Two notifier lists: the "policy" list is involved in the
116  * validation process for a new CPU frequency policy; the
117  * "transition" list for kernel code that needs to handle
118  * changes to devices when the CPU clock speed changes.
119  * The mutex locks both lists.
120  */
121 static BLOCKING_NOTIFIER_HEAD(cpufreq_policy_notifier_list);
122 static struct srcu_notifier_head cpufreq_transition_notifier_list;
123
124 static bool init_cpufreq_transition_notifier_list_called;
125 static int __init init_cpufreq_transition_notifier_list(void)
126 {
127         srcu_init_notifier_head(&cpufreq_transition_notifier_list);
128         init_cpufreq_transition_notifier_list_called = true;
129         return 0;
130 }
131 pure_initcall(init_cpufreq_transition_notifier_list);
132
133 static LIST_HEAD(cpufreq_governor_list);
134 static DEFINE_MUTEX(cpufreq_governor_mutex);
135
136 struct cpufreq_policy *cpufreq_cpu_get(unsigned int cpu)
137 {
138         struct cpufreq_policy *data;
139         unsigned long flags;
140
141         if (cpu >= nr_cpu_ids)
142                 goto err_out;
143
144         /* get the cpufreq driver */
145         spin_lock_irqsave(&cpufreq_driver_lock, flags);
146
147         if (!cpufreq_driver)
148                 goto err_out_unlock;
149
150         if (!try_module_get(cpufreq_driver->owner))
151                 goto err_out_unlock;
152
153
154         /* get the CPU */
155         data = per_cpu(cpufreq_cpu_data, cpu);
156
157         if (!data)
158                 goto err_out_put_module;
159
160         if (!kobject_get(&data->kobj))
161                 goto err_out_put_module;
162
163         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
164         return data;
165
166 err_out_put_module:
167         module_put(cpufreq_driver->owner);
168 err_out_unlock:
169         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
170 err_out:
171         return NULL;
172 }
173 EXPORT_SYMBOL_GPL(cpufreq_cpu_get);
174
175
176 void cpufreq_cpu_put(struct cpufreq_policy *data)
177 {
178         kobject_put(&data->kobj);
179         module_put(cpufreq_driver->owner);
180 }
181 EXPORT_SYMBOL_GPL(cpufreq_cpu_put);
182
183
184 /*********************************************************************
185  *                     UNIFIED DEBUG HELPERS                         *
186  *********************************************************************/
187 #ifdef CONFIG_CPU_FREQ_DEBUG
188
189 /* what part(s) of the CPUfreq subsystem are debugged? */
190 static unsigned int debug;
191
192 /* is the debug output ratelimit'ed using printk_ratelimit? User can
193  * set or modify this value.
194  */
195 static unsigned int debug_ratelimit = 1;
196
197 /* is the printk_ratelimit'ing enabled? It's enabled after a successful
198  * loading of a cpufreq driver, temporarily disabled when a new policy
199  * is set, and disabled upon cpufreq driver removal
200  */
201 static unsigned int disable_ratelimit = 1;
202 static DEFINE_SPINLOCK(disable_ratelimit_lock);
203
204 static void cpufreq_debug_enable_ratelimit(void)
205 {
206         unsigned long flags;
207
208         spin_lock_irqsave(&disable_ratelimit_lock, flags);
209         if (disable_ratelimit)
210                 disable_ratelimit--;
211         spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
212 }
213
214 static void cpufreq_debug_disable_ratelimit(void)
215 {
216         unsigned long flags;
217
218         spin_lock_irqsave(&disable_ratelimit_lock, flags);
219         disable_ratelimit++;
220         spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
221 }
222
223 void cpufreq_debug_printk(unsigned int type, const char *prefix,
224                         const char *fmt, ...)
225 {
226         char s[256];
227         va_list args;
228         unsigned int len;
229         unsigned long flags;
230
231         WARN_ON(!prefix);
232         if (type & debug) {
233                 spin_lock_irqsave(&disable_ratelimit_lock, flags);
234                 if (!disable_ratelimit && debug_ratelimit
235                                         && !printk_ratelimit()) {
236                         spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
237                         return;
238                 }
239                 spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
240
241                 len = snprintf(s, 256, KERN_DEBUG "%s: ", prefix);
242
243                 va_start(args, fmt);
244                 len += vsnprintf(&s[len], (256 - len), fmt, args);
245                 va_end(args);
246
247                 printk(s);
248
249                 WARN_ON(len < 5);
250         }
251 }
252 EXPORT_SYMBOL(cpufreq_debug_printk);
253
254
255 module_param(debug, uint, 0644);
256 MODULE_PARM_DESC(debug, "CPUfreq debugging: add 1 to debug core,"
257                         " 2 to debug drivers, and 4 to debug governors.");
258
259 module_param(debug_ratelimit, uint, 0644);
260 MODULE_PARM_DESC(debug_ratelimit, "CPUfreq debugging:"
261                                         " set to 0 to disable ratelimiting.");
262
263 #else /* !CONFIG_CPU_FREQ_DEBUG */
264
265 static inline void cpufreq_debug_enable_ratelimit(void) { return; }
266 static inline void cpufreq_debug_disable_ratelimit(void) { return; }
267
268 #endif /* CONFIG_CPU_FREQ_DEBUG */
269
270
271 /*********************************************************************
272  *            EXTERNALLY AFFECTING FREQUENCY CHANGES                 *
273  *********************************************************************/
274
275 /**
276  * adjust_jiffies - adjust the system "loops_per_jiffy"
277  *
278  * This function alters the system "loops_per_jiffy" for the clock
279  * speed change. Note that loops_per_jiffy cannot be updated on SMP
280  * systems as each CPU might be scaled differently. So, use the arch
281  * per-CPU loops_per_jiffy value wherever possible.
282  */
283 #ifndef CONFIG_SMP
284 static unsigned long l_p_j_ref;
285 static unsigned int  l_p_j_ref_freq;
286
287 static void adjust_jiffies(unsigned long val, struct cpufreq_freqs *ci)
288 {
289         if (ci->flags & CPUFREQ_CONST_LOOPS)
290                 return;
291
292         if (!l_p_j_ref_freq) {
293                 l_p_j_ref = loops_per_jiffy;
294                 l_p_j_ref_freq = ci->old;
295                 dprintk("saving %lu as reference value for loops_per_jiffy; "
296                         "freq is %u kHz\n", l_p_j_ref, l_p_j_ref_freq);
297         }
298         if ((val == CPUFREQ_PRECHANGE  && ci->old < ci->new) ||
299             (val == CPUFREQ_POSTCHANGE && ci->old > ci->new) ||
300             (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
301                 loops_per_jiffy = cpufreq_scale(l_p_j_ref, l_p_j_ref_freq,
302                                                                 ci->new);
303                 dprintk("scaling loops_per_jiffy to %lu "
304                         "for frequency %u kHz\n", loops_per_jiffy, ci->new);
305         }
306 }
307 #else
308 static inline void adjust_jiffies(unsigned long val, struct cpufreq_freqs *ci)
309 {
310         return;
311 }
312 #endif
313
314
315 /**
316  * cpufreq_notify_transition - call notifier chain and adjust_jiffies
317  * on frequency transition.
318  *
319  * This function calls the transition notifiers and the "adjust_jiffies"
320  * function. It is called twice on all CPU frequency changes that have
321  * external effects.
322  */
323 void cpufreq_notify_transition(struct cpufreq_freqs *freqs, unsigned int state)
324 {
325         struct cpufreq_policy *policy;
326
327         BUG_ON(irqs_disabled());
328
329         freqs->flags = cpufreq_driver->flags;
330         dprintk("notification %u of frequency transition to %u kHz\n",
331                 state, freqs->new);
332
333         policy = per_cpu(cpufreq_cpu_data, freqs->cpu);
334         switch (state) {
335
336         case CPUFREQ_PRECHANGE:
337                 /* detect if the driver reported a value as "old frequency"
338                  * which is not equal to what the cpufreq core thinks is
339                  * "old frequency".
340                  */
341                 if (!(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
342                         if ((policy) && (policy->cpu == freqs->cpu) &&
343                             (policy->cur) && (policy->cur != freqs->old)) {
344                                 dprintk("Warning: CPU frequency is"
345                                         " %u, cpufreq assumed %u kHz.\n",
346                                         freqs->old, policy->cur);
347                                 freqs->old = policy->cur;
348                         }
349                 }
350                 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
351                                 CPUFREQ_PRECHANGE, freqs);
352                 adjust_jiffies(CPUFREQ_PRECHANGE, freqs);
353                 break;
354
355         case CPUFREQ_POSTCHANGE:
356                 adjust_jiffies(CPUFREQ_POSTCHANGE, freqs);
357                 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
358                                 CPUFREQ_POSTCHANGE, freqs);
359                 if (likely(policy) && likely(policy->cpu == freqs->cpu))
360                         policy->cur = freqs->new;
361                 break;
362         }
363 }
364 EXPORT_SYMBOL_GPL(cpufreq_notify_transition);
365
366
367
368 /*********************************************************************
369  *                          SYSFS INTERFACE                          *
370  *********************************************************************/
371
372 static struct cpufreq_governor *__find_governor(const char *str_governor)
373 {
374         struct cpufreq_governor *t;
375
376         list_for_each_entry(t, &cpufreq_governor_list, governor_list)
377                 if (!strnicmp(str_governor, t->name, CPUFREQ_NAME_LEN))
378                         return t;
379
380         return NULL;
381 }
382
383 /**
384  * cpufreq_parse_governor - parse a governor string
385  */
386 static int cpufreq_parse_governor(char *str_governor, unsigned int *policy,
387                                 struct cpufreq_governor **governor)
388 {
389         int err = -EINVAL;
390
391         if (!cpufreq_driver)
392                 goto out;
393
394         if (cpufreq_driver->setpolicy) {
395                 if (!strnicmp(str_governor, "performance", CPUFREQ_NAME_LEN)) {
396                         *policy = CPUFREQ_POLICY_PERFORMANCE;
397                         err = 0;
398                 } else if (!strnicmp(str_governor, "powersave",
399                                                 CPUFREQ_NAME_LEN)) {
400                         *policy = CPUFREQ_POLICY_POWERSAVE;
401                         err = 0;
402                 }
403         } else if (cpufreq_driver->target) {
404                 struct cpufreq_governor *t;
405
406                 mutex_lock(&cpufreq_governor_mutex);
407
408                 t = __find_governor(str_governor);
409
410                 if (t == NULL) {
411                         char *name = kasprintf(GFP_KERNEL, "cpufreq_%s",
412                                                                 str_governor);
413
414                         if (name) {
415                                 int ret;
416
417                                 mutex_unlock(&cpufreq_governor_mutex);
418                                 ret = request_module("%s", name);
419                                 mutex_lock(&cpufreq_governor_mutex);
420
421                                 if (ret == 0)
422                                         t = __find_governor(str_governor);
423                         }
424
425                         kfree(name);
426                 }
427
428                 if (t != NULL) {
429                         *governor = t;
430                         err = 0;
431                 }
432
433                 mutex_unlock(&cpufreq_governor_mutex);
434         }
435 out:
436         return err;
437 }
438
439
440 /**
441  * cpufreq_per_cpu_attr_read() / show_##file_name() -
442  * print out cpufreq information
443  *
444  * Write out information from cpufreq_driver->policy[cpu]; object must be
445  * "unsigned int".
446  */
447
448 #define show_one(file_name, object)                     \
449 static ssize_t show_##file_name                         \
450 (struct cpufreq_policy *policy, char *buf)              \
451 {                                                       \
452         return sprintf(buf, "%u\n", policy->object);    \
453 }
454
455 show_one(cpuinfo_min_freq, cpuinfo.min_freq);
456 show_one(cpuinfo_max_freq, cpuinfo.max_freq);
457 show_one(cpuinfo_transition_latency, cpuinfo.transition_latency);
458 show_one(scaling_min_freq, min);
459 show_one(scaling_max_freq, max);
460 show_one(scaling_cur_freq, cur);
461
462 static int __cpufreq_set_policy(struct cpufreq_policy *data,
463                                 struct cpufreq_policy *policy);
464
465 /**
466  * cpufreq_per_cpu_attr_write() / store_##file_name() - sysfs write access
467  */
468 #define store_one(file_name, object)                    \
469 static ssize_t store_##file_name                                        \
470 (struct cpufreq_policy *policy, const char *buf, size_t count)          \
471 {                                                                       \
472         unsigned int ret = -EINVAL;                                     \
473         struct cpufreq_policy new_policy;                               \
474                                                                         \
475         ret = cpufreq_get_policy(&new_policy, policy->cpu);             \
476         if (ret)                                                        \
477                 return -EINVAL;                                         \
478                                                                         \
479         ret = sscanf(buf, "%u", &new_policy.object);                    \
480         if (ret != 1)                                                   \
481                 return -EINVAL;                                         \
482                                                                         \
483         ret = __cpufreq_set_policy(policy, &new_policy);                \
484         policy->user_policy.object = policy->object;                    \
485                                                                         \
486         return ret ? ret : count;                                       \
487 }
488
489 store_one(scaling_min_freq, min);
490 store_one(scaling_max_freq, max);
491
492 /**
493  * show_cpuinfo_cur_freq - current CPU frequency as detected by hardware
494  */
495 static ssize_t show_cpuinfo_cur_freq(struct cpufreq_policy *policy,
496                                         char *buf)
497 {
498         unsigned int cur_freq = __cpufreq_get(policy->cpu);
499         if (!cur_freq)
500                 return sprintf(buf, "<unknown>");
501         return sprintf(buf, "%u\n", cur_freq);
502 }
503
504
505 /**
506  * show_scaling_governor - show the current policy for the specified CPU
507  */
508 static ssize_t show_scaling_governor(struct cpufreq_policy *policy, char *buf)
509 {
510         if (policy->policy == CPUFREQ_POLICY_POWERSAVE)
511                 return sprintf(buf, "powersave\n");
512         else if (policy->policy == CPUFREQ_POLICY_PERFORMANCE)
513                 return sprintf(buf, "performance\n");
514         else if (policy->governor)
515                 return scnprintf(buf, CPUFREQ_NAME_LEN, "%s\n",
516                                 policy->governor->name);
517         return -EINVAL;
518 }
519
520
521 /**
522  * store_scaling_governor - store policy for the specified CPU
523  */
524 static ssize_t store_scaling_governor(struct cpufreq_policy *policy,
525                                         const char *buf, size_t count)
526 {
527         unsigned int ret = -EINVAL;
528         char    str_governor[16];
529         struct cpufreq_policy new_policy;
530
531         ret = cpufreq_get_policy(&new_policy, policy->cpu);
532         if (ret)
533                 return ret;
534
535         ret = sscanf(buf, "%15s", str_governor);
536         if (ret != 1)
537                 return -EINVAL;
538
539         if (cpufreq_parse_governor(str_governor, &new_policy.policy,
540                                                 &new_policy.governor))
541                 return -EINVAL;
542
543         /* Do not use cpufreq_set_policy here or the user_policy.max
544            will be wrongly overridden */
545         ret = __cpufreq_set_policy(policy, &new_policy);
546
547         policy->user_policy.policy = policy->policy;
548         policy->user_policy.governor = policy->governor;
549
550         if (ret)
551                 return ret;
552         else
553                 return count;
554 }
555
556 /**
557  * show_scaling_driver - show the cpufreq driver currently loaded
558  */
559 static ssize_t show_scaling_driver(struct cpufreq_policy *policy, char *buf)
560 {
561         return scnprintf(buf, CPUFREQ_NAME_LEN, "%s\n", cpufreq_driver->name);
562 }
563
564 /**
565  * show_scaling_available_governors - show the available CPUfreq governors
566  */
567 static ssize_t show_scaling_available_governors(struct cpufreq_policy *policy,
568                                                 char *buf)
569 {
570         ssize_t i = 0;
571         struct cpufreq_governor *t;
572
573         if (!cpufreq_driver->target) {
574                 i += sprintf(buf, "performance powersave");
575                 goto out;
576         }
577
578         list_for_each_entry(t, &cpufreq_governor_list, governor_list) {
579                 if (i >= (ssize_t) ((PAGE_SIZE / sizeof(char))
580                     - (CPUFREQ_NAME_LEN + 2)))
581                         goto out;
582                 i += scnprintf(&buf[i], CPUFREQ_NAME_LEN, "%s ", t->name);
583         }
584 out:
585         i += sprintf(&buf[i], "\n");
586         return i;
587 }
588
589 static ssize_t show_cpus(const struct cpumask *mask, char *buf)
590 {
591         ssize_t i = 0;
592         unsigned int cpu;
593
594         for_each_cpu(cpu, mask) {
595                 if (i)
596                         i += scnprintf(&buf[i], (PAGE_SIZE - i - 2), " ");
597                 i += scnprintf(&buf[i], (PAGE_SIZE - i - 2), "%u", cpu);
598                 if (i >= (PAGE_SIZE - 5))
599                         break;
600         }
601         i += sprintf(&buf[i], "\n");
602         return i;
603 }
604
605 /**
606  * show_related_cpus - show the CPUs affected by each transition even if
607  * hw coordination is in use
608  */
609 static ssize_t show_related_cpus(struct cpufreq_policy *policy, char *buf)
610 {
611         if (cpumask_empty(policy->related_cpus))
612                 return show_cpus(policy->cpus, buf);
613         return show_cpus(policy->related_cpus, buf);
614 }
615
616 /**
617  * show_affected_cpus - show the CPUs affected by each transition
618  */
619 static ssize_t show_affected_cpus(struct cpufreq_policy *policy, char *buf)
620 {
621         return show_cpus(policy->cpus, buf);
622 }
623
624 static ssize_t store_scaling_setspeed(struct cpufreq_policy *policy,
625                                         const char *buf, size_t count)
626 {
627         unsigned int freq = 0;
628         unsigned int ret;
629
630         if (!policy->governor || !policy->governor->store_setspeed)
631                 return -EINVAL;
632
633         ret = sscanf(buf, "%u", &freq);
634         if (ret != 1)
635                 return -EINVAL;
636
637         policy->governor->store_setspeed(policy, freq);
638
639         return count;
640 }
641
642 static ssize_t show_scaling_setspeed(struct cpufreq_policy *policy, char *buf)
643 {
644         if (!policy->governor || !policy->governor->show_setspeed)
645                 return sprintf(buf, "<unsupported>\n");
646
647         return policy->governor->show_setspeed(policy, buf);
648 }
649
650 /**
651  * show_scaling_driver - show the current cpufreq HW/BIOS limitation
652  */
653 static ssize_t show_bios_limit(struct cpufreq_policy *policy, char *buf)
654 {
655         unsigned int limit;
656         int ret;
657         if (cpufreq_driver->bios_limit) {
658                 ret = cpufreq_driver->bios_limit(policy->cpu, &limit);
659                 if (!ret)
660                         return sprintf(buf, "%u\n", limit);
661         }
662         return sprintf(buf, "%u\n", policy->cpuinfo.max_freq);
663 }
664
665 #define define_one_ro(_name) \
666 static struct freq_attr _name = \
667 __ATTR(_name, 0444, show_##_name, NULL)
668
669 #define define_one_ro0400(_name) \
670 static struct freq_attr _name = \
671 __ATTR(_name, 0400, show_##_name, NULL)
672
673 #define define_one_rw(_name) \
674 static struct freq_attr _name = \
675 __ATTR(_name, 0644, show_##_name, store_##_name)
676
677 define_one_ro0400(cpuinfo_cur_freq);
678 define_one_ro(cpuinfo_min_freq);
679 define_one_ro(cpuinfo_max_freq);
680 define_one_ro(cpuinfo_transition_latency);
681 define_one_ro(scaling_available_governors);
682 define_one_ro(scaling_driver);
683 define_one_ro(scaling_cur_freq);
684 define_one_ro(bios_limit);
685 define_one_ro(related_cpus);
686 define_one_ro(affected_cpus);
687 define_one_rw(scaling_min_freq);
688 define_one_rw(scaling_max_freq);
689 define_one_rw(scaling_governor);
690 define_one_rw(scaling_setspeed);
691
692 static struct attribute *default_attrs[] = {
693         &cpuinfo_min_freq.attr,
694         &cpuinfo_max_freq.attr,
695         &cpuinfo_transition_latency.attr,
696         &scaling_min_freq.attr,
697         &scaling_max_freq.attr,
698         &affected_cpus.attr,
699         &related_cpus.attr,
700         &scaling_governor.attr,
701         &scaling_driver.attr,
702         &scaling_available_governors.attr,
703         &scaling_setspeed.attr,
704         NULL
705 };
706
707 struct kobject *cpufreq_global_kobject;
708 EXPORT_SYMBOL(cpufreq_global_kobject);
709
710 #define to_policy(k) container_of(k, struct cpufreq_policy, kobj)
711 #define to_attr(a) container_of(a, struct freq_attr, attr)
712
713 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
714 {
715         struct cpufreq_policy *policy = to_policy(kobj);
716         struct freq_attr *fattr = to_attr(attr);
717         ssize_t ret = -EINVAL;
718         policy = cpufreq_cpu_get(policy->cpu);
719         if (!policy)
720                 goto no_policy;
721
722         if (lock_policy_rwsem_read(policy->cpu) < 0)
723                 goto fail;
724
725         if (fattr->show)
726                 ret = fattr->show(policy, buf);
727         else
728                 ret = -EIO;
729
730         unlock_policy_rwsem_read(policy->cpu);
731 fail:
732         cpufreq_cpu_put(policy);
733 no_policy:
734         return ret;
735 }
736
737 static ssize_t store(struct kobject *kobj, struct attribute *attr,
738                      const char *buf, size_t count)
739 {
740         struct cpufreq_policy *policy = to_policy(kobj);
741         struct freq_attr *fattr = to_attr(attr);
742         ssize_t ret = -EINVAL;
743         policy = cpufreq_cpu_get(policy->cpu);
744         if (!policy)
745                 goto no_policy;
746
747         if (lock_policy_rwsem_write(policy->cpu) < 0)
748                 goto fail;
749
750         if (fattr->store)
751                 ret = fattr->store(policy, buf, count);
752         else
753                 ret = -EIO;
754
755         unlock_policy_rwsem_write(policy->cpu);
756 fail:
757         cpufreq_cpu_put(policy);
758 no_policy:
759         return ret;
760 }
761
762 static void cpufreq_sysfs_release(struct kobject *kobj)
763 {
764         struct cpufreq_policy *policy = to_policy(kobj);
765         dprintk("last reference is dropped\n");
766         complete(&policy->kobj_unregister);
767 }
768
769 static struct sysfs_ops sysfs_ops = {
770         .show   = show,
771         .store  = store,
772 };
773
774 static struct kobj_type ktype_cpufreq = {
775         .sysfs_ops      = &sysfs_ops,
776         .default_attrs  = default_attrs,
777         .release        = cpufreq_sysfs_release,
778 };
779
780 /*
781  * Returns:
782  *   Negative: Failure
783  *   0:        Success
784  *   Positive: When we have a managed CPU and the sysfs got symlinked
785  */
786 static int cpufreq_add_dev_policy(unsigned int cpu,
787                                   struct cpufreq_policy *policy,
788                                   struct sys_device *sys_dev)
789 {
790         int ret = 0;
791 #ifdef CONFIG_SMP
792         unsigned long flags;
793         unsigned int j;
794 #ifdef CONFIG_HOTPLUG_CPU
795         struct cpufreq_governor *gov;
796
797         gov = __find_governor(per_cpu(cpufreq_cpu_governor, cpu));
798         if (gov) {
799                 policy->governor = gov;
800                 dprintk("Restoring governor %s for cpu %d\n",
801                        policy->governor->name, cpu);
802         }
803 #endif
804
805         for_each_cpu(j, policy->cpus) {
806                 struct cpufreq_policy *managed_policy;
807
808                 if (cpu == j)
809                         continue;
810
811                 /* Check for existing affected CPUs.
812                  * They may not be aware of it due to CPU Hotplug.
813                  * cpufreq_cpu_put is called when the device is removed
814                  * in __cpufreq_remove_dev()
815                  */
816                 managed_policy = cpufreq_cpu_get(j);
817                 if (unlikely(managed_policy)) {
818
819                         /* Set proper policy_cpu */
820                         unlock_policy_rwsem_write(cpu);
821                         per_cpu(cpufreq_policy_cpu, cpu) = managed_policy->cpu;
822
823                         if (lock_policy_rwsem_write(cpu) < 0) {
824                                 /* Should not go through policy unlock path */
825                                 if (cpufreq_driver->exit)
826                                         cpufreq_driver->exit(policy);
827                                 cpufreq_cpu_put(managed_policy);
828                                 return -EBUSY;
829                         }
830
831                         spin_lock_irqsave(&cpufreq_driver_lock, flags);
832                         cpumask_copy(managed_policy->cpus, policy->cpus);
833                         per_cpu(cpufreq_cpu_data, cpu) = managed_policy;
834                         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
835
836                         dprintk("CPU already managed, adding link\n");
837                         ret = sysfs_create_link(&sys_dev->kobj,
838                                                 &managed_policy->kobj,
839                                                 "cpufreq");
840                         if (ret)
841                                 cpufreq_cpu_put(managed_policy);
842                         /*
843                          * Success. We only needed to be added to the mask.
844                          * Call driver->exit() because only the cpu parent of
845                          * the kobj needed to call init().
846                          */
847                         if (cpufreq_driver->exit)
848                                 cpufreq_driver->exit(policy);
849
850                         if (!ret)
851                                 return 1;
852                         else
853                                 return ret;
854                 }
855         }
856 #endif
857         return ret;
858 }
859
860
861 /* symlink affected CPUs */
862 static int cpufreq_add_dev_symlink(unsigned int cpu,
863                                    struct cpufreq_policy *policy)
864 {
865         unsigned int j;
866         int ret = 0;
867
868         for_each_cpu(j, policy->cpus) {
869                 struct cpufreq_policy *managed_policy;
870                 struct sys_device *cpu_sys_dev;
871
872                 if (j == cpu)
873                         continue;
874                 if (!cpu_online(j))
875                         continue;
876
877                 dprintk("CPU %u already managed, adding link\n", j);
878                 managed_policy = cpufreq_cpu_get(cpu);
879                 cpu_sys_dev = get_cpu_sysdev(j);
880                 ret = sysfs_create_link(&cpu_sys_dev->kobj, &policy->kobj,
881                                         "cpufreq");
882                 if (ret) {
883                         cpufreq_cpu_put(managed_policy);
884                         return ret;
885                 }
886         }
887         return ret;
888 }
889
890 static int cpufreq_add_dev_interface(unsigned int cpu,
891                                      struct cpufreq_policy *policy,
892                                      struct sys_device *sys_dev)
893 {
894         struct cpufreq_policy new_policy;
895         struct freq_attr **drv_attr;
896         unsigned long flags;
897         int ret = 0;
898         unsigned int j;
899
900         /* prepare interface data */
901         ret = kobject_init_and_add(&policy->kobj, &ktype_cpufreq,
902                                    &sys_dev->kobj, "cpufreq");
903         if (ret)
904                 return ret;
905
906         /* set up files for this cpu device */
907         drv_attr = cpufreq_driver->attr;
908         while ((drv_attr) && (*drv_attr)) {
909                 ret = sysfs_create_file(&policy->kobj, &((*drv_attr)->attr));
910                 if (ret)
911                         goto err_out_kobj_put;
912                 drv_attr++;
913         }
914         if (cpufreq_driver->get) {
915                 ret = sysfs_create_file(&policy->kobj, &cpuinfo_cur_freq.attr);
916                 if (ret)
917                         goto err_out_kobj_put;
918         }
919         if (cpufreq_driver->target) {
920                 ret = sysfs_create_file(&policy->kobj, &scaling_cur_freq.attr);
921                 if (ret)
922                         goto err_out_kobj_put;
923         }
924         if (cpufreq_driver->bios_limit) {
925                 ret = sysfs_create_file(&policy->kobj, &bios_limit.attr);
926                 if (ret)
927                         goto err_out_kobj_put;
928         }
929
930         spin_lock_irqsave(&cpufreq_driver_lock, flags);
931         for_each_cpu(j, policy->cpus) {
932         if (!cpu_online(j))
933                 continue;
934                 per_cpu(cpufreq_cpu_data, j) = policy;
935                 per_cpu(cpufreq_policy_cpu, j) = policy->cpu;
936         }
937         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
938
939         ret = cpufreq_add_dev_symlink(cpu, policy);
940         if (ret)
941                 goto err_out_kobj_put;
942
943         memcpy(&new_policy, policy, sizeof(struct cpufreq_policy));
944         /* assure that the starting sequence is run in __cpufreq_set_policy */
945         policy->governor = NULL;
946
947         /* set default policy */
948         ret = __cpufreq_set_policy(policy, &new_policy);
949         policy->user_policy.policy = policy->policy;
950         policy->user_policy.governor = policy->governor;
951
952         if (ret) {
953                 dprintk("setting policy failed\n");
954                 if (cpufreq_driver->exit)
955                         cpufreq_driver->exit(policy);
956         }
957         return ret;
958
959 err_out_kobj_put:
960         kobject_put(&policy->kobj);
961         wait_for_completion(&policy->kobj_unregister);
962         return ret;
963 }
964
965
966 /**
967  * cpufreq_add_dev - add a CPU device
968  *
969  * Adds the cpufreq interface for a CPU device.
970  *
971  * The Oracle says: try running cpufreq registration/unregistration concurrently
972  * with with cpu hotplugging and all hell will break loose. Tried to clean this
973  * mess up, but more thorough testing is needed. - Mathieu
974  */
975 static int cpufreq_add_dev(struct sys_device *sys_dev)
976 {
977         unsigned int cpu = sys_dev->id;
978         int ret = 0, found = 0;
979         struct cpufreq_policy *policy;
980         unsigned long flags;
981         unsigned int j;
982 #ifdef CONFIG_HOTPLUG_CPU
983         int sibling;
984 #endif
985
986         if (cpu_is_offline(cpu))
987                 return 0;
988
989         cpufreq_debug_disable_ratelimit();
990         dprintk("adding CPU %u\n", cpu);
991
992 #ifdef CONFIG_SMP
993         /* check whether a different CPU already registered this
994          * CPU because it is in the same boat. */
995         policy = cpufreq_cpu_get(cpu);
996         if (unlikely(policy)) {
997                 cpufreq_cpu_put(policy);
998                 cpufreq_debug_enable_ratelimit();
999                 return 0;
1000         }
1001 #endif
1002
1003         if (!try_module_get(cpufreq_driver->owner)) {
1004                 ret = -EINVAL;
1005                 goto module_out;
1006         }
1007
1008         ret = -ENOMEM;
1009         policy = kzalloc(sizeof(struct cpufreq_policy), GFP_KERNEL);
1010         if (!policy)
1011                 goto nomem_out;
1012
1013         if (!alloc_cpumask_var(&policy->cpus, GFP_KERNEL))
1014                 goto err_free_policy;
1015
1016         if (!zalloc_cpumask_var(&policy->related_cpus, GFP_KERNEL))
1017                 goto err_free_cpumask;
1018
1019         policy->cpu = cpu;
1020         cpumask_copy(policy->cpus, cpumask_of(cpu));
1021
1022         /* Initially set CPU itself as the policy_cpu */
1023         per_cpu(cpufreq_policy_cpu, cpu) = cpu;
1024         ret = (lock_policy_rwsem_write(cpu) < 0);
1025         WARN_ON(ret);
1026
1027         init_completion(&policy->kobj_unregister);
1028         INIT_WORK(&policy->update, handle_update);
1029
1030         /* Set governor before ->init, so that driver could check it */
1031 #ifdef CONFIG_HOTPLUG_CPU
1032         for_each_online_cpu(sibling) {
1033                 struct cpufreq_policy *cp = per_cpu(cpufreq_cpu_data, sibling);
1034                 if (cp && cp->governor &&
1035                     (cpumask_test_cpu(cpu, cp->related_cpus))) {
1036                         policy->governor = cp->governor;
1037                         found = 1;
1038                         break;
1039                 }
1040         }
1041 #endif
1042         if (!found)
1043                 policy->governor = CPUFREQ_DEFAULT_GOVERNOR;
1044         /* call driver. From then on the cpufreq must be able
1045          * to accept all calls to ->verify and ->setpolicy for this CPU
1046          */
1047         ret = cpufreq_driver->init(policy);
1048         if (ret) {
1049                 dprintk("initialization failed\n");
1050                 goto err_unlock_policy;
1051         }
1052         policy->user_policy.min = policy->min;
1053         policy->user_policy.max = policy->max;
1054
1055         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1056                                      CPUFREQ_START, policy);
1057
1058         ret = cpufreq_add_dev_policy(cpu, policy, sys_dev);
1059         if (ret) {
1060                 if (ret > 0)
1061                         /* This is a managed cpu, symlink created,
1062                            exit with 0 */
1063                         ret = 0;
1064                 goto err_unlock_policy;
1065         }
1066
1067         ret = cpufreq_add_dev_interface(cpu, policy, sys_dev);
1068         if (ret)
1069                 goto err_out_unregister;
1070
1071         unlock_policy_rwsem_write(cpu);
1072
1073         kobject_uevent(&policy->kobj, KOBJ_ADD);
1074         module_put(cpufreq_driver->owner);
1075         dprintk("initialization complete\n");
1076         cpufreq_debug_enable_ratelimit();
1077
1078         return 0;
1079
1080
1081 err_out_unregister:
1082         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1083         for_each_cpu(j, policy->cpus)
1084                 per_cpu(cpufreq_cpu_data, j) = NULL;
1085         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1086
1087         kobject_put(&policy->kobj);
1088         wait_for_completion(&policy->kobj_unregister);
1089
1090 err_unlock_policy:
1091         unlock_policy_rwsem_write(cpu);
1092 err_free_cpumask:
1093         free_cpumask_var(policy->cpus);
1094 err_free_policy:
1095         kfree(policy);
1096 nomem_out:
1097         module_put(cpufreq_driver->owner);
1098 module_out:
1099         cpufreq_debug_enable_ratelimit();
1100         return ret;
1101 }
1102
1103
1104 /**
1105  * __cpufreq_remove_dev - remove a CPU device
1106  *
1107  * Removes the cpufreq interface for a CPU device.
1108  * Caller should already have policy_rwsem in write mode for this CPU.
1109  * This routine frees the rwsem before returning.
1110  */
1111 static int __cpufreq_remove_dev(struct sys_device *sys_dev)
1112 {
1113         unsigned int cpu = sys_dev->id;
1114         unsigned long flags;
1115         struct cpufreq_policy *data;
1116 #ifdef CONFIG_SMP
1117         struct sys_device *cpu_sys_dev;
1118         unsigned int j;
1119 #endif
1120
1121         cpufreq_debug_disable_ratelimit();
1122         dprintk("unregistering CPU %u\n", cpu);
1123
1124         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1125         data = per_cpu(cpufreq_cpu_data, cpu);
1126
1127         if (!data) {
1128                 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1129                 cpufreq_debug_enable_ratelimit();
1130                 unlock_policy_rwsem_write(cpu);
1131                 return -EINVAL;
1132         }
1133         per_cpu(cpufreq_cpu_data, cpu) = NULL;
1134
1135
1136 #ifdef CONFIG_SMP
1137         /* if this isn't the CPU which is the parent of the kobj, we
1138          * only need to unlink, put and exit
1139          */
1140         if (unlikely(cpu != data->cpu)) {
1141                 dprintk("removing link\n");
1142                 cpumask_clear_cpu(cpu, data->cpus);
1143                 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1144                 sysfs_remove_link(&sys_dev->kobj, "cpufreq");
1145                 cpufreq_cpu_put(data);
1146                 cpufreq_debug_enable_ratelimit();
1147                 unlock_policy_rwsem_write(cpu);
1148                 return 0;
1149         }
1150 #endif
1151
1152 #ifdef CONFIG_SMP
1153
1154 #ifdef CONFIG_HOTPLUG_CPU
1155         strncpy(per_cpu(cpufreq_cpu_governor, cpu), data->governor->name,
1156                         CPUFREQ_NAME_LEN);
1157 #endif
1158
1159         /* if we have other CPUs still registered, we need to unlink them,
1160          * or else wait_for_completion below will lock up. Clean the
1161          * per_cpu(cpufreq_cpu_data) while holding the lock, and remove
1162          * the sysfs links afterwards.
1163          */
1164         if (unlikely(cpumask_weight(data->cpus) > 1)) {
1165                 for_each_cpu(j, data->cpus) {
1166                         if (j == cpu)
1167                                 continue;
1168                         per_cpu(cpufreq_cpu_data, j) = NULL;
1169                 }
1170         }
1171
1172         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1173
1174         if (unlikely(cpumask_weight(data->cpus) > 1)) {
1175                 for_each_cpu(j, data->cpus) {
1176                         if (j == cpu)
1177                                 continue;
1178                         dprintk("removing link for cpu %u\n", j);
1179 #ifdef CONFIG_HOTPLUG_CPU
1180                         strncpy(per_cpu(cpufreq_cpu_governor, j),
1181                                 data->governor->name, CPUFREQ_NAME_LEN);
1182 #endif
1183                         cpu_sys_dev = get_cpu_sysdev(j);
1184                         sysfs_remove_link(&cpu_sys_dev->kobj, "cpufreq");
1185                         cpufreq_cpu_put(data);
1186                 }
1187         }
1188 #else
1189         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1190 #endif
1191
1192         if (cpufreq_driver->target)
1193                 __cpufreq_governor(data, CPUFREQ_GOV_STOP);
1194
1195         kobject_put(&data->kobj);
1196
1197         /* we need to make sure that the underlying kobj is actually
1198          * not referenced anymore by anybody before we proceed with
1199          * unloading.
1200          */
1201         dprintk("waiting for dropping of refcount\n");
1202         wait_for_completion(&data->kobj_unregister);
1203         dprintk("wait complete\n");
1204
1205         if (cpufreq_driver->exit)
1206                 cpufreq_driver->exit(data);
1207
1208         unlock_policy_rwsem_write(cpu);
1209
1210         free_cpumask_var(data->related_cpus);
1211         free_cpumask_var(data->cpus);
1212         kfree(data);
1213         per_cpu(cpufreq_cpu_data, cpu) = NULL;
1214
1215         cpufreq_debug_enable_ratelimit();
1216         return 0;
1217 }
1218
1219
1220 static int cpufreq_remove_dev(struct sys_device *sys_dev)
1221 {
1222         unsigned int cpu = sys_dev->id;
1223         int retval;
1224
1225         if (cpu_is_offline(cpu))
1226                 return 0;
1227
1228         if (unlikely(lock_policy_rwsem_write(cpu)))
1229                 BUG();
1230
1231         retval = __cpufreq_remove_dev(sys_dev);
1232         return retval;
1233 }
1234
1235
1236 static void handle_update(struct work_struct *work)
1237 {
1238         struct cpufreq_policy *policy =
1239                 container_of(work, struct cpufreq_policy, update);
1240         unsigned int cpu = policy->cpu;
1241         dprintk("handle_update for cpu %u called\n", cpu);
1242         cpufreq_update_policy(cpu);
1243 }
1244
1245 /**
1246  *      cpufreq_out_of_sync - If actual and saved CPU frequency differs, we're in deep trouble.
1247  *      @cpu: cpu number
1248  *      @old_freq: CPU frequency the kernel thinks the CPU runs at
1249  *      @new_freq: CPU frequency the CPU actually runs at
1250  *
1251  *      We adjust to current frequency first, and need to clean up later.
1252  *      So either call to cpufreq_update_policy() or schedule handle_update()).
1253  */
1254 static void cpufreq_out_of_sync(unsigned int cpu, unsigned int old_freq,
1255                                 unsigned int new_freq)
1256 {
1257         struct cpufreq_freqs freqs;
1258
1259         dprintk("Warning: CPU frequency out of sync: cpufreq and timing "
1260                "core thinks of %u, is %u kHz.\n", old_freq, new_freq);
1261
1262         freqs.cpu = cpu;
1263         freqs.old = old_freq;
1264         freqs.new = new_freq;
1265         cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1266         cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1267 }
1268
1269
1270 /**
1271  * cpufreq_quick_get - get the CPU frequency (in kHz) from policy->cur
1272  * @cpu: CPU number
1273  *
1274  * This is the last known freq, without actually getting it from the driver.
1275  * Return value will be same as what is shown in scaling_cur_freq in sysfs.
1276  */
1277 unsigned int cpufreq_quick_get(unsigned int cpu)
1278 {
1279         struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
1280         unsigned int ret_freq = 0;
1281
1282         if (policy) {
1283                 ret_freq = policy->cur;
1284                 cpufreq_cpu_put(policy);
1285         }
1286
1287         return ret_freq;
1288 }
1289 EXPORT_SYMBOL(cpufreq_quick_get);
1290
1291
1292 static unsigned int __cpufreq_get(unsigned int cpu)
1293 {
1294         struct cpufreq_policy *policy = per_cpu(cpufreq_cpu_data, cpu);
1295         unsigned int ret_freq = 0;
1296
1297         if (!cpufreq_driver->get)
1298                 return ret_freq;
1299
1300         ret_freq = cpufreq_driver->get(cpu);
1301
1302         if (ret_freq && policy->cur &&
1303                 !(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
1304                 /* verify no discrepancy between actual and
1305                                         saved value exists */
1306                 if (unlikely(ret_freq != policy->cur)) {
1307                         cpufreq_out_of_sync(cpu, policy->cur, ret_freq);
1308                         schedule_work(&policy->update);
1309                 }
1310         }
1311
1312         return ret_freq;
1313 }
1314
1315 /**
1316  * cpufreq_get - get the current CPU frequency (in kHz)
1317  * @cpu: CPU number
1318  *
1319  * Get the CPU current (static) CPU frequency
1320  */
1321 unsigned int cpufreq_get(unsigned int cpu)
1322 {
1323         unsigned int ret_freq = 0;
1324         struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
1325
1326         if (!policy)
1327                 goto out;
1328
1329         if (unlikely(lock_policy_rwsem_read(cpu)))
1330                 goto out_policy;
1331
1332         ret_freq = __cpufreq_get(cpu);
1333
1334         unlock_policy_rwsem_read(cpu);
1335
1336 out_policy:
1337         cpufreq_cpu_put(policy);
1338 out:
1339         return ret_freq;
1340 }
1341 EXPORT_SYMBOL(cpufreq_get);
1342
1343
1344 /**
1345  *      cpufreq_suspend - let the low level driver prepare for suspend
1346  */
1347
1348 static int cpufreq_suspend(struct sys_device *sysdev, pm_message_t pmsg)
1349 {
1350         int ret = 0;
1351
1352         int cpu = sysdev->id;
1353         struct cpufreq_policy *cpu_policy;
1354
1355         dprintk("suspending cpu %u\n", cpu);
1356
1357         if (!cpu_online(cpu))
1358                 return 0;
1359
1360         /* we may be lax here as interrupts are off. Nonetheless
1361          * we need to grab the correct cpu policy, as to check
1362          * whether we really run on this CPU.
1363          */
1364
1365         cpu_policy = cpufreq_cpu_get(cpu);
1366         if (!cpu_policy)
1367                 return -EINVAL;
1368
1369         /* only handle each CPU group once */
1370         if (unlikely(cpu_policy->cpu != cpu))
1371                 goto out;
1372
1373         if (cpufreq_driver->suspend) {
1374                 ret = cpufreq_driver->suspend(cpu_policy, pmsg);
1375                 if (ret)
1376                         printk(KERN_ERR "cpufreq: suspend failed in ->suspend "
1377                                         "step on CPU %u\n", cpu_policy->cpu);
1378         }
1379
1380 out:
1381         cpufreq_cpu_put(cpu_policy);
1382         return ret;
1383 }
1384
1385 /**
1386  *      cpufreq_resume -  restore proper CPU frequency handling after resume
1387  *
1388  *      1.) resume CPUfreq hardware support (cpufreq_driver->resume())
1389  *      2.) schedule call cpufreq_update_policy() ASAP as interrupts are
1390  *          restored. It will verify that the current freq is in sync with
1391  *          what we believe it to be. This is a bit later than when it
1392  *          should be, but nonethteless it's better than calling
1393  *          cpufreq_driver->get() here which might re-enable interrupts...
1394  */
1395 static int cpufreq_resume(struct sys_device *sysdev)
1396 {
1397         int ret = 0;
1398
1399         int cpu = sysdev->id;
1400         struct cpufreq_policy *cpu_policy;
1401
1402         dprintk("resuming cpu %u\n", cpu);
1403
1404         if (!cpu_online(cpu))
1405                 return 0;
1406
1407         /* we may be lax here as interrupts are off. Nonetheless
1408          * we need to grab the correct cpu policy, as to check
1409          * whether we really run on this CPU.
1410          */
1411
1412         cpu_policy = cpufreq_cpu_get(cpu);
1413         if (!cpu_policy)
1414                 return -EINVAL;
1415
1416         /* only handle each CPU group once */
1417         if (unlikely(cpu_policy->cpu != cpu))
1418                 goto fail;
1419
1420         if (cpufreq_driver->resume) {
1421                 ret = cpufreq_driver->resume(cpu_policy);
1422                 if (ret) {
1423                         printk(KERN_ERR "cpufreq: resume failed in ->resume "
1424                                         "step on CPU %u\n", cpu_policy->cpu);
1425                         goto fail;
1426                 }
1427         }
1428
1429         schedule_work(&cpu_policy->update);
1430
1431 fail:
1432         cpufreq_cpu_put(cpu_policy);
1433         return ret;
1434 }
1435
1436 static struct sysdev_driver cpufreq_sysdev_driver = {
1437         .add            = cpufreq_add_dev,
1438         .remove         = cpufreq_remove_dev,
1439         .suspend        = cpufreq_suspend,
1440         .resume         = cpufreq_resume,
1441 };
1442
1443
1444 /*********************************************************************
1445  *                     NOTIFIER LISTS INTERFACE                      *
1446  *********************************************************************/
1447
1448 /**
1449  *      cpufreq_register_notifier - register a driver with cpufreq
1450  *      @nb: notifier function to register
1451  *      @list: CPUFREQ_TRANSITION_NOTIFIER or CPUFREQ_POLICY_NOTIFIER
1452  *
1453  *      Add a driver to one of two lists: either a list of drivers that
1454  *      are notified about clock rate changes (once before and once after
1455  *      the transition), or a list of drivers that are notified about
1456  *      changes in cpufreq policy.
1457  *
1458  *      This function may sleep, and has the same return conditions as
1459  *      blocking_notifier_chain_register.
1460  */
1461 int cpufreq_register_notifier(struct notifier_block *nb, unsigned int list)
1462 {
1463         int ret;
1464
1465         WARN_ON(!init_cpufreq_transition_notifier_list_called);
1466
1467         switch (list) {
1468         case CPUFREQ_TRANSITION_NOTIFIER:
1469                 ret = srcu_notifier_chain_register(
1470                                 &cpufreq_transition_notifier_list, nb);
1471                 break;
1472         case CPUFREQ_POLICY_NOTIFIER:
1473                 ret = blocking_notifier_chain_register(
1474                                 &cpufreq_policy_notifier_list, nb);
1475                 break;
1476         default:
1477                 ret = -EINVAL;
1478         }
1479
1480         return ret;
1481 }
1482 EXPORT_SYMBOL(cpufreq_register_notifier);
1483
1484
1485 /**
1486  *      cpufreq_unregister_notifier - unregister a driver with cpufreq
1487  *      @nb: notifier block to be unregistered
1488  *      @list: CPUFREQ_TRANSITION_NOTIFIER or CPUFREQ_POLICY_NOTIFIER
1489  *
1490  *      Remove a driver from the CPU frequency notifier list.
1491  *
1492  *      This function may sleep, and has the same return conditions as
1493  *      blocking_notifier_chain_unregister.
1494  */
1495 int cpufreq_unregister_notifier(struct notifier_block *nb, unsigned int list)
1496 {
1497         int ret;
1498
1499         switch (list) {
1500         case CPUFREQ_TRANSITION_NOTIFIER:
1501                 ret = srcu_notifier_chain_unregister(
1502                                 &cpufreq_transition_notifier_list, nb);
1503                 break;
1504         case CPUFREQ_POLICY_NOTIFIER:
1505                 ret = blocking_notifier_chain_unregister(
1506                                 &cpufreq_policy_notifier_list, nb);
1507                 break;
1508         default:
1509                 ret = -EINVAL;
1510         }
1511
1512         return ret;
1513 }
1514 EXPORT_SYMBOL(cpufreq_unregister_notifier);
1515
1516
1517 /*********************************************************************
1518  *                              GOVERNORS                            *
1519  *********************************************************************/
1520
1521
1522 int __cpufreq_driver_target(struct cpufreq_policy *policy,
1523                             unsigned int target_freq,
1524                             unsigned int relation)
1525 {
1526         int retval = -EINVAL;
1527
1528         dprintk("target for CPU %u: %u kHz, relation %u\n", policy->cpu,
1529                 target_freq, relation);
1530         if (cpu_online(policy->cpu) && cpufreq_driver->target)
1531                 retval = cpufreq_driver->target(policy, target_freq, relation);
1532
1533         return retval;
1534 }
1535 EXPORT_SYMBOL_GPL(__cpufreq_driver_target);
1536
1537 int cpufreq_driver_target(struct cpufreq_policy *policy,
1538                           unsigned int target_freq,
1539                           unsigned int relation)
1540 {
1541         int ret = -EINVAL;
1542
1543         policy = cpufreq_cpu_get(policy->cpu);
1544         if (!policy)
1545                 goto no_policy;
1546
1547         if (unlikely(lock_policy_rwsem_write(policy->cpu)))
1548                 goto fail;
1549
1550         ret = __cpufreq_driver_target(policy, target_freq, relation);
1551
1552         unlock_policy_rwsem_write(policy->cpu);
1553
1554 fail:
1555         cpufreq_cpu_put(policy);
1556 no_policy:
1557         return ret;
1558 }
1559 EXPORT_SYMBOL_GPL(cpufreq_driver_target);
1560
1561 int __cpufreq_driver_getavg(struct cpufreq_policy *policy, unsigned int cpu)
1562 {
1563         int ret = 0;
1564
1565         policy = cpufreq_cpu_get(policy->cpu);
1566         if (!policy)
1567                 return -EINVAL;
1568
1569         if (cpu_online(cpu) && cpufreq_driver->getavg)
1570                 ret = cpufreq_driver->getavg(policy, cpu);
1571
1572         cpufreq_cpu_put(policy);
1573         return ret;
1574 }
1575 EXPORT_SYMBOL_GPL(__cpufreq_driver_getavg);
1576
1577 /*
1578  * when "event" is CPUFREQ_GOV_LIMITS
1579  */
1580
1581 static int __cpufreq_governor(struct cpufreq_policy *policy,
1582                                         unsigned int event)
1583 {
1584         int ret;
1585
1586         /* Only must be defined when default governor is known to have latency
1587            restrictions, like e.g. conservative or ondemand.
1588            That this is the case is already ensured in Kconfig
1589         */
1590 #ifdef CONFIG_CPU_FREQ_GOV_PERFORMANCE
1591         struct cpufreq_governor *gov = &cpufreq_gov_performance;
1592 #else
1593         struct cpufreq_governor *gov = NULL;
1594 #endif
1595
1596         if (policy->governor->max_transition_latency &&
1597             policy->cpuinfo.transition_latency >
1598             policy->governor->max_transition_latency) {
1599                 if (!gov)
1600                         return -EINVAL;
1601                 else {
1602                         printk(KERN_WARNING "%s governor failed, too long"
1603                                " transition latency of HW, fallback"
1604                                " to %s governor\n",
1605                                policy->governor->name,
1606                                gov->name);
1607                         policy->governor = gov;
1608                 }
1609         }
1610
1611         if (!try_module_get(policy->governor->owner))
1612                 return -EINVAL;
1613
1614         dprintk("__cpufreq_governor for CPU %u, event %u\n",
1615                                                 policy->cpu, event);
1616         ret = policy->governor->governor(policy, event);
1617
1618         /* we keep one module reference alive for
1619                         each CPU governed by this CPU */
1620         if ((event != CPUFREQ_GOV_START) || ret)
1621                 module_put(policy->governor->owner);
1622         if ((event == CPUFREQ_GOV_STOP) && !ret)
1623                 module_put(policy->governor->owner);
1624
1625         return ret;
1626 }
1627
1628
1629 int cpufreq_register_governor(struct cpufreq_governor *governor)
1630 {
1631         int err;
1632
1633         if (!governor)
1634                 return -EINVAL;
1635
1636         mutex_lock(&cpufreq_governor_mutex);
1637
1638         err = -EBUSY;
1639         if (__find_governor(governor->name) == NULL) {
1640                 err = 0;
1641                 list_add(&governor->governor_list, &cpufreq_governor_list);
1642         }
1643
1644         mutex_unlock(&cpufreq_governor_mutex);
1645         return err;
1646 }
1647 EXPORT_SYMBOL_GPL(cpufreq_register_governor);
1648
1649
1650 void cpufreq_unregister_governor(struct cpufreq_governor *governor)
1651 {
1652 #ifdef CONFIG_HOTPLUG_CPU
1653         int cpu;
1654 #endif
1655
1656         if (!governor)
1657                 return;
1658
1659 #ifdef CONFIG_HOTPLUG_CPU
1660         for_each_present_cpu(cpu) {
1661                 if (cpu_online(cpu))
1662                         continue;
1663                 if (!strcmp(per_cpu(cpufreq_cpu_governor, cpu), governor->name))
1664                         strcpy(per_cpu(cpufreq_cpu_governor, cpu), "\0");
1665         }
1666 #endif
1667
1668         mutex_lock(&cpufreq_governor_mutex);
1669         list_del(&governor->governor_list);
1670         mutex_unlock(&cpufreq_governor_mutex);
1671         return;
1672 }
1673 EXPORT_SYMBOL_GPL(cpufreq_unregister_governor);
1674
1675
1676
1677 /*********************************************************************
1678  *                          POLICY INTERFACE                         *
1679  *********************************************************************/
1680
1681 /**
1682  * cpufreq_get_policy - get the current cpufreq_policy
1683  * @policy: struct cpufreq_policy into which the current cpufreq_policy
1684  *      is written
1685  *
1686  * Reads the current cpufreq policy.
1687  */
1688 int cpufreq_get_policy(struct cpufreq_policy *policy, unsigned int cpu)
1689 {
1690         struct cpufreq_policy *cpu_policy;
1691         if (!policy)
1692                 return -EINVAL;
1693
1694         cpu_policy = cpufreq_cpu_get(cpu);
1695         if (!cpu_policy)
1696                 return -EINVAL;
1697
1698         memcpy(policy, cpu_policy, sizeof(struct cpufreq_policy));
1699
1700         cpufreq_cpu_put(cpu_policy);
1701         return 0;
1702 }
1703 EXPORT_SYMBOL(cpufreq_get_policy);
1704
1705
1706 /*
1707  * data   : current policy.
1708  * policy : policy to be set.
1709  */
1710 static int __cpufreq_set_policy(struct cpufreq_policy *data,
1711                                 struct cpufreq_policy *policy)
1712 {
1713         int ret = 0;
1714
1715         cpufreq_debug_disable_ratelimit();
1716         dprintk("setting new policy for CPU %u: %u - %u kHz\n", policy->cpu,
1717                 policy->min, policy->max);
1718
1719         memcpy(&policy->cpuinfo, &data->cpuinfo,
1720                                 sizeof(struct cpufreq_cpuinfo));
1721
1722         if (policy->min > data->max || policy->max < data->min) {
1723                 ret = -EINVAL;
1724                 goto error_out;
1725         }
1726
1727         /* verify the cpu speed can be set within this limit */
1728         ret = cpufreq_driver->verify(policy);
1729         if (ret)
1730                 goto error_out;
1731
1732         /* adjust if necessary - all reasons */
1733         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1734                         CPUFREQ_ADJUST, policy);
1735
1736         /* adjust if necessary - hardware incompatibility*/
1737         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1738                         CPUFREQ_INCOMPATIBLE, policy);
1739
1740         /* verify the cpu speed can be set within this limit,
1741            which might be different to the first one */
1742         ret = cpufreq_driver->verify(policy);
1743         if (ret)
1744                 goto error_out;
1745
1746         /* notification of the new policy */
1747         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1748                         CPUFREQ_NOTIFY, policy);
1749
1750         data->min = policy->min;
1751         data->max = policy->max;
1752
1753         dprintk("new min and max freqs are %u - %u kHz\n",
1754                                         data->min, data->max);
1755
1756         if (cpufreq_driver->setpolicy) {
1757                 data->policy = policy->policy;
1758                 dprintk("setting range\n");
1759                 ret = cpufreq_driver->setpolicy(policy);
1760         } else {
1761                 if (policy->governor != data->governor) {
1762                         /* save old, working values */
1763                         struct cpufreq_governor *old_gov = data->governor;
1764
1765                         dprintk("governor switch\n");
1766
1767                         /* end old governor */
1768                         if (data->governor) {
1769                                 /*
1770                                  * Need to release the rwsem around governor
1771                                  * stop due to lock dependency between
1772                                  * cancel_delayed_work_sync and the read lock
1773                                  * taken in the delayed work handler.
1774                                  */
1775                                 unlock_policy_rwsem_write(data->cpu);
1776                                 __cpufreq_governor(data, CPUFREQ_GOV_STOP);
1777                                 lock_policy_rwsem_write(data->cpu);
1778                         }
1779
1780                         /* start new governor */
1781                         data->governor = policy->governor;
1782                         if (__cpufreq_governor(data, CPUFREQ_GOV_START)) {
1783                                 /* new governor failed, so re-start old one */
1784                                 dprintk("starting governor %s failed\n",
1785                                                         data->governor->name);
1786                                 if (old_gov) {
1787                                         data->governor = old_gov;
1788                                         __cpufreq_governor(data,
1789                                                            CPUFREQ_GOV_START);
1790                                 }
1791                                 ret = -EINVAL;
1792                                 goto error_out;
1793                         }
1794                         /* might be a policy change, too, so fall through */
1795                 }
1796                 dprintk("governor: change or update limits\n");
1797                 __cpufreq_governor(data, CPUFREQ_GOV_LIMITS);
1798         }
1799
1800 error_out:
1801         cpufreq_debug_enable_ratelimit();
1802         return ret;
1803 }
1804
1805 /**
1806  *      cpufreq_update_policy - re-evaluate an existing cpufreq policy
1807  *      @cpu: CPU which shall be re-evaluated
1808  *
1809  *      Usefull for policy notifiers which have different necessities
1810  *      at different times.
1811  */
1812 int cpufreq_update_policy(unsigned int cpu)
1813 {
1814         struct cpufreq_policy *data = cpufreq_cpu_get(cpu);
1815         struct cpufreq_policy policy;
1816         int ret;
1817
1818         if (!data) {
1819                 ret = -ENODEV;
1820                 goto no_policy;
1821         }
1822
1823         if (unlikely(lock_policy_rwsem_write(cpu))) {
1824                 ret = -EINVAL;
1825                 goto fail;
1826         }
1827
1828         dprintk("updating policy for CPU %u\n", cpu);
1829         memcpy(&policy, data, sizeof(struct cpufreq_policy));
1830         policy.min = data->user_policy.min;
1831         policy.max = data->user_policy.max;
1832         policy.policy = data->user_policy.policy;
1833         policy.governor = data->user_policy.governor;
1834
1835         /* BIOS might change freq behind our back
1836           -> ask driver for current freq and notify governors about a change */
1837         if (cpufreq_driver->get) {
1838                 policy.cur = cpufreq_driver->get(cpu);
1839                 if (!data->cur) {
1840                         dprintk("Driver did not initialize current freq");
1841                         data->cur = policy.cur;
1842                 } else {
1843                         if (data->cur != policy.cur)
1844                                 cpufreq_out_of_sync(cpu, data->cur,
1845                                                                 policy.cur);
1846                 }
1847         }
1848
1849         ret = __cpufreq_set_policy(data, &policy);
1850
1851         unlock_policy_rwsem_write(cpu);
1852
1853 fail:
1854         cpufreq_cpu_put(data);
1855 no_policy:
1856         return ret;
1857 }
1858 EXPORT_SYMBOL(cpufreq_update_policy);
1859
1860 static int __cpuinit cpufreq_cpu_callback(struct notifier_block *nfb,
1861                                         unsigned long action, void *hcpu)
1862 {
1863         unsigned int cpu = (unsigned long)hcpu;
1864         struct sys_device *sys_dev;
1865
1866         sys_dev = get_cpu_sysdev(cpu);
1867         if (sys_dev) {
1868                 switch (action) {
1869                 case CPU_ONLINE:
1870                 case CPU_ONLINE_FROZEN:
1871                         cpufreq_add_dev(sys_dev);
1872                         break;
1873                 case CPU_DOWN_PREPARE:
1874                 case CPU_DOWN_PREPARE_FROZEN:
1875                         if (unlikely(lock_policy_rwsem_write(cpu)))
1876                                 BUG();
1877
1878                         __cpufreq_remove_dev(sys_dev);
1879                         break;
1880                 case CPU_DOWN_FAILED:
1881                 case CPU_DOWN_FAILED_FROZEN:
1882                         cpufreq_add_dev(sys_dev);
1883                         break;
1884                 }
1885         }
1886         return NOTIFY_OK;
1887 }
1888
1889 static struct notifier_block __refdata cpufreq_cpu_notifier =
1890 {
1891     .notifier_call = cpufreq_cpu_callback,
1892 };
1893
1894 /*********************************************************************
1895  *               REGISTER / UNREGISTER CPUFREQ DRIVER                *
1896  *********************************************************************/
1897
1898 /**
1899  * cpufreq_register_driver - register a CPU Frequency driver
1900  * @driver_data: A struct cpufreq_driver containing the values#
1901  * submitted by the CPU Frequency driver.
1902  *
1903  *   Registers a CPU Frequency driver to this core code. This code
1904  * returns zero on success, -EBUSY when another driver got here first
1905  * (and isn't unregistered in the meantime).
1906  *
1907  */
1908 int cpufreq_register_driver(struct cpufreq_driver *driver_data)
1909 {
1910         unsigned long flags;
1911         int ret;
1912
1913         if (!driver_data || !driver_data->verify || !driver_data->init ||
1914             ((!driver_data->setpolicy) && (!driver_data->target)))
1915                 return -EINVAL;
1916
1917         dprintk("trying to register driver %s\n", driver_data->name);
1918
1919         if (driver_data->setpolicy)
1920                 driver_data->flags |= CPUFREQ_CONST_LOOPS;
1921
1922         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1923         if (cpufreq_driver) {
1924                 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1925                 return -EBUSY;
1926         }
1927         cpufreq_driver = driver_data;
1928         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1929
1930         ret = sysdev_driver_register(&cpu_sysdev_class,
1931                                         &cpufreq_sysdev_driver);
1932
1933         if ((!ret) && !(cpufreq_driver->flags & CPUFREQ_STICKY)) {
1934                 int i;
1935                 ret = -ENODEV;
1936
1937                 /* check for at least one working CPU */
1938                 for (i = 0; i < nr_cpu_ids; i++)
1939                         if (cpu_possible(i) && per_cpu(cpufreq_cpu_data, i)) {
1940                                 ret = 0;
1941                                 break;
1942                         }
1943
1944                 /* if all ->init() calls failed, unregister */
1945                 if (ret) {
1946                         dprintk("no CPU initialized for driver %s\n",
1947                                                         driver_data->name);
1948                         sysdev_driver_unregister(&cpu_sysdev_class,
1949                                                 &cpufreq_sysdev_driver);
1950
1951                         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1952                         cpufreq_driver = NULL;
1953                         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1954                 }
1955         }
1956
1957         if (!ret) {
1958                 register_hotcpu_notifier(&cpufreq_cpu_notifier);
1959                 dprintk("driver %s up and running\n", driver_data->name);
1960                 cpufreq_debug_enable_ratelimit();
1961         }
1962
1963         return ret;
1964 }
1965 EXPORT_SYMBOL_GPL(cpufreq_register_driver);
1966
1967
1968 /**
1969  * cpufreq_unregister_driver - unregister the current CPUFreq driver
1970  *
1971  *    Unregister the current CPUFreq driver. Only call this if you have
1972  * the right to do so, i.e. if you have succeeded in initialising before!
1973  * Returns zero if successful, and -EINVAL if the cpufreq_driver is
1974  * currently not initialised.
1975  */
1976 int cpufreq_unregister_driver(struct cpufreq_driver *driver)
1977 {
1978         unsigned long flags;
1979
1980         cpufreq_debug_disable_ratelimit();
1981
1982         if (!cpufreq_driver || (driver != cpufreq_driver)) {
1983                 cpufreq_debug_enable_ratelimit();
1984                 return -EINVAL;
1985         }
1986
1987         dprintk("unregistering driver %s\n", driver->name);
1988
1989         sysdev_driver_unregister(&cpu_sysdev_class, &cpufreq_sysdev_driver);
1990         unregister_hotcpu_notifier(&cpufreq_cpu_notifier);
1991
1992         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1993         cpufreq_driver = NULL;
1994         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1995
1996         return 0;
1997 }
1998 EXPORT_SYMBOL_GPL(cpufreq_unregister_driver);
1999
2000 static int __init cpufreq_core_init(void)
2001 {
2002         int cpu;
2003
2004         for_each_possible_cpu(cpu) {
2005                 per_cpu(cpufreq_policy_cpu, cpu) = -1;
2006                 init_rwsem(&per_cpu(cpu_policy_rwsem, cpu));
2007         }
2008
2009         cpufreq_global_kobject = kobject_create_and_add("cpufreq",
2010                                                 &cpu_sysdev_class.kset.kobj);
2011         BUG_ON(!cpufreq_global_kobject);
2012
2013         return 0;
2014 }
2015 core_initcall(cpufreq_core_init);