2 * linux/drivers/cpufreq/cpufreq.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2002 - 2003 Dominik Brodowski <linux@brodo.de>
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
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.
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>
32 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_CORE, \
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.
40 static struct cpufreq_driver *cpufreq_driver;
41 static struct cpufreq_policy *cpufreq_cpu_data[NR_CPUS];
42 static DEFINE_SPINLOCK(cpufreq_driver_lock);
44 /* internal prototypes */
45 static int __cpufreq_governor(struct cpufreq_policy *policy, unsigned int event);
46 static void handle_update(struct work_struct *work);
49 * Two notifier lists: the "policy" list is involved in the
50 * validation process for a new CPU frequency policy; the
51 * "transition" list for kernel code that needs to handle
52 * changes to devices when the CPU clock speed changes.
53 * The mutex locks both lists.
55 static BLOCKING_NOTIFIER_HEAD(cpufreq_policy_notifier_list);
56 static struct srcu_notifier_head cpufreq_transition_notifier_list;
58 static int __init init_cpufreq_transition_notifier_list(void)
60 srcu_init_notifier_head(&cpufreq_transition_notifier_list);
63 pure_initcall(init_cpufreq_transition_notifier_list);
65 static LIST_HEAD(cpufreq_governor_list);
66 static DEFINE_MUTEX (cpufreq_governor_mutex);
68 struct cpufreq_policy *cpufreq_cpu_get(unsigned int cpu)
70 struct cpufreq_policy *data;
76 /* get the cpufreq driver */
77 spin_lock_irqsave(&cpufreq_driver_lock, flags);
82 if (!try_module_get(cpufreq_driver->owner))
87 data = cpufreq_cpu_data[cpu];
90 goto err_out_put_module;
92 if (!kobject_get(&data->kobj))
93 goto err_out_put_module;
95 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
99 module_put(cpufreq_driver->owner);
101 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
105 EXPORT_SYMBOL_GPL(cpufreq_cpu_get);
108 void cpufreq_cpu_put(struct cpufreq_policy *data)
110 kobject_put(&data->kobj);
111 module_put(cpufreq_driver->owner);
113 EXPORT_SYMBOL_GPL(cpufreq_cpu_put);
116 /*********************************************************************
117 * UNIFIED DEBUG HELPERS *
118 *********************************************************************/
119 #ifdef CONFIG_CPU_FREQ_DEBUG
121 /* what part(s) of the CPUfreq subsystem are debugged? */
122 static unsigned int debug;
124 /* is the debug output ratelimit'ed using printk_ratelimit? User can
125 * set or modify this value.
127 static unsigned int debug_ratelimit = 1;
129 /* is the printk_ratelimit'ing enabled? It's enabled after a successful
130 * loading of a cpufreq driver, temporarily disabled when a new policy
131 * is set, and disabled upon cpufreq driver removal
133 static unsigned int disable_ratelimit = 1;
134 static DEFINE_SPINLOCK(disable_ratelimit_lock);
136 static void cpufreq_debug_enable_ratelimit(void)
140 spin_lock_irqsave(&disable_ratelimit_lock, flags);
141 if (disable_ratelimit)
143 spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
146 static void cpufreq_debug_disable_ratelimit(void)
150 spin_lock_irqsave(&disable_ratelimit_lock, flags);
152 spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
155 void cpufreq_debug_printk(unsigned int type, const char *prefix,
156 const char *fmt, ...)
165 spin_lock_irqsave(&disable_ratelimit_lock, flags);
166 if (!disable_ratelimit && debug_ratelimit
167 && !printk_ratelimit()) {
168 spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
171 spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
173 len = snprintf(s, 256, KERN_DEBUG "%s: ", prefix);
176 len += vsnprintf(&s[len], (256 - len), fmt, args);
184 EXPORT_SYMBOL(cpufreq_debug_printk);
187 module_param(debug, uint, 0644);
188 MODULE_PARM_DESC(debug, "CPUfreq debugging: add 1 to debug core,"
189 " 2 to debug drivers, and 4 to debug governors.");
191 module_param(debug_ratelimit, uint, 0644);
192 MODULE_PARM_DESC(debug_ratelimit, "CPUfreq debugging:"
193 " set to 0 to disable ratelimiting.");
195 #else /* !CONFIG_CPU_FREQ_DEBUG */
197 static inline void cpufreq_debug_enable_ratelimit(void) { return; }
198 static inline void cpufreq_debug_disable_ratelimit(void) { return; }
200 #endif /* CONFIG_CPU_FREQ_DEBUG */
203 /*********************************************************************
204 * EXTERNALLY AFFECTING FREQUENCY CHANGES *
205 *********************************************************************/
208 * adjust_jiffies - adjust the system "loops_per_jiffy"
210 * This function alters the system "loops_per_jiffy" for the clock
211 * speed change. Note that loops_per_jiffy cannot be updated on SMP
212 * systems as each CPU might be scaled differently. So, use the arch
213 * per-CPU loops_per_jiffy value wherever possible.
216 static unsigned long l_p_j_ref;
217 static unsigned int l_p_j_ref_freq;
219 static void adjust_jiffies(unsigned long val, struct cpufreq_freqs *ci)
221 if (ci->flags & CPUFREQ_CONST_LOOPS)
224 if (!l_p_j_ref_freq) {
225 l_p_j_ref = loops_per_jiffy;
226 l_p_j_ref_freq = ci->old;
227 dprintk("saving %lu as reference value for loops_per_jiffy;"
228 "freq is %u kHz\n", l_p_j_ref, l_p_j_ref_freq);
230 if ((val == CPUFREQ_PRECHANGE && ci->old < ci->new) ||
231 (val == CPUFREQ_POSTCHANGE && ci->old > ci->new) ||
232 (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
233 loops_per_jiffy = cpufreq_scale(l_p_j_ref, l_p_j_ref_freq,
235 dprintk("scaling loops_per_jiffy to %lu"
236 "for frequency %u kHz\n", loops_per_jiffy, ci->new);
240 static inline void adjust_jiffies(unsigned long val, struct cpufreq_freqs *ci)
248 * cpufreq_notify_transition - call notifier chain and adjust_jiffies
249 * on frequency transition.
251 * This function calls the transition notifiers and the "adjust_jiffies"
252 * function. It is called twice on all CPU frequency changes that have
255 void cpufreq_notify_transition(struct cpufreq_freqs *freqs, unsigned int state)
257 struct cpufreq_policy *policy;
259 BUG_ON(irqs_disabled());
261 freqs->flags = cpufreq_driver->flags;
262 dprintk("notification %u of frequency transition to %u kHz\n",
265 policy = cpufreq_cpu_data[freqs->cpu];
268 case CPUFREQ_PRECHANGE:
269 /* detect if the driver reported a value as "old frequency"
270 * which is not equal to what the cpufreq core thinks is
273 if (!(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
274 if ((policy) && (policy->cpu == freqs->cpu) &&
275 (policy->cur) && (policy->cur != freqs->old)) {
276 dprintk("Warning: CPU frequency is"
277 " %u, cpufreq assumed %u kHz.\n",
278 freqs->old, policy->cur);
279 freqs->old = policy->cur;
282 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
283 CPUFREQ_PRECHANGE, freqs);
284 adjust_jiffies(CPUFREQ_PRECHANGE, freqs);
287 case CPUFREQ_POSTCHANGE:
288 adjust_jiffies(CPUFREQ_POSTCHANGE, freqs);
289 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
290 CPUFREQ_POSTCHANGE, freqs);
291 if (likely(policy) && likely(policy->cpu == freqs->cpu))
292 policy->cur = freqs->new;
296 EXPORT_SYMBOL_GPL(cpufreq_notify_transition);
300 /*********************************************************************
302 *********************************************************************/
304 static struct cpufreq_governor *__find_governor(const char *str_governor)
306 struct cpufreq_governor *t;
308 list_for_each_entry(t, &cpufreq_governor_list, governor_list)
309 if (!strnicmp(str_governor,t->name,CPUFREQ_NAME_LEN))
316 * cpufreq_parse_governor - parse a governor string
318 static int cpufreq_parse_governor (char *str_governor, unsigned int *policy,
319 struct cpufreq_governor **governor)
326 if (cpufreq_driver->setpolicy) {
327 if (!strnicmp(str_governor, "performance", CPUFREQ_NAME_LEN)) {
328 *policy = CPUFREQ_POLICY_PERFORMANCE;
330 } else if (!strnicmp(str_governor, "powersave",
332 *policy = CPUFREQ_POLICY_POWERSAVE;
335 } else if (cpufreq_driver->target) {
336 struct cpufreq_governor *t;
338 mutex_lock(&cpufreq_governor_mutex);
340 t = __find_governor(str_governor);
343 char *name = kasprintf(GFP_KERNEL, "cpufreq_%s",
349 mutex_unlock(&cpufreq_governor_mutex);
350 ret = request_module(name);
351 mutex_lock(&cpufreq_governor_mutex);
354 t = __find_governor(str_governor);
365 mutex_unlock(&cpufreq_governor_mutex);
372 /* drivers/base/cpu.c */
373 extern struct sysdev_class cpu_sysdev_class;
377 * cpufreq_per_cpu_attr_read() / show_##file_name() -
378 * print out cpufreq information
380 * Write out information from cpufreq_driver->policy[cpu]; object must be
384 #define show_one(file_name, object) \
385 static ssize_t show_##file_name \
386 (struct cpufreq_policy * policy, char *buf) \
388 return sprintf (buf, "%u\n", policy->object); \
391 show_one(cpuinfo_min_freq, cpuinfo.min_freq);
392 show_one(cpuinfo_max_freq, cpuinfo.max_freq);
393 show_one(scaling_min_freq, min);
394 show_one(scaling_max_freq, max);
395 show_one(scaling_cur_freq, cur);
397 static int __cpufreq_set_policy(struct cpufreq_policy *data,
398 struct cpufreq_policy *policy);
401 * cpufreq_per_cpu_attr_write() / store_##file_name() - sysfs write access
403 #define store_one(file_name, object) \
404 static ssize_t store_##file_name \
405 (struct cpufreq_policy * policy, const char *buf, size_t count) \
407 unsigned int ret = -EINVAL; \
408 struct cpufreq_policy new_policy; \
410 ret = cpufreq_get_policy(&new_policy, policy->cpu); \
414 ret = sscanf (buf, "%u", &new_policy.object); \
418 mutex_lock(&policy->lock); \
419 ret = __cpufreq_set_policy(policy, &new_policy); \
420 policy->user_policy.object = policy->object; \
421 mutex_unlock(&policy->lock); \
423 return ret ? ret : count; \
426 store_one(scaling_min_freq,min);
427 store_one(scaling_max_freq,max);
430 * show_cpuinfo_cur_freq - current CPU frequency as detected by hardware
432 static ssize_t show_cpuinfo_cur_freq (struct cpufreq_policy * policy,
435 unsigned int cur_freq = cpufreq_get(policy->cpu);
437 return sprintf(buf, "<unknown>");
438 return sprintf(buf, "%u\n", cur_freq);
443 * show_scaling_governor - show the current policy for the specified CPU
445 static ssize_t show_scaling_governor (struct cpufreq_policy * policy,
448 if(policy->policy == CPUFREQ_POLICY_POWERSAVE)
449 return sprintf(buf, "powersave\n");
450 else if (policy->policy == CPUFREQ_POLICY_PERFORMANCE)
451 return sprintf(buf, "performance\n");
452 else if (policy->governor)
453 return scnprintf(buf, CPUFREQ_NAME_LEN, "%s\n", policy->governor->name);
459 * store_scaling_governor - store policy for the specified CPU
461 static ssize_t store_scaling_governor (struct cpufreq_policy * policy,
462 const char *buf, size_t count)
464 unsigned int ret = -EINVAL;
465 char str_governor[16];
466 struct cpufreq_policy new_policy;
468 ret = cpufreq_get_policy(&new_policy, policy->cpu);
472 ret = sscanf (buf, "%15s", str_governor);
476 if (cpufreq_parse_governor(str_governor, &new_policy.policy,
477 &new_policy.governor))
480 /* Do not use cpufreq_set_policy here or the user_policy.max
481 will be wrongly overridden */
482 mutex_lock(&policy->lock);
483 ret = __cpufreq_set_policy(policy, &new_policy);
485 policy->user_policy.policy = policy->policy;
486 policy->user_policy.governor = policy->governor;
487 mutex_unlock(&policy->lock);
496 * show_scaling_driver - show the cpufreq driver currently loaded
498 static ssize_t show_scaling_driver (struct cpufreq_policy * policy, char *buf)
500 return scnprintf(buf, CPUFREQ_NAME_LEN, "%s\n", cpufreq_driver->name);
504 * show_scaling_available_governors - show the available CPUfreq governors
506 static ssize_t show_scaling_available_governors (struct cpufreq_policy *policy,
510 struct cpufreq_governor *t;
512 if (!cpufreq_driver->target) {
513 i += sprintf(buf, "performance powersave");
517 list_for_each_entry(t, &cpufreq_governor_list, governor_list) {
518 if (i >= (ssize_t) ((PAGE_SIZE / sizeof(char)) - (CPUFREQ_NAME_LEN + 2)))
520 i += scnprintf(&buf[i], CPUFREQ_NAME_LEN, "%s ", t->name);
523 i += sprintf(&buf[i], "\n");
527 * show_affected_cpus - show the CPUs affected by each transition
529 static ssize_t show_affected_cpus (struct cpufreq_policy * policy, char *buf)
534 for_each_cpu_mask(cpu, policy->cpus) {
536 i += scnprintf(&buf[i], (PAGE_SIZE - i - 2), " ");
537 i += scnprintf(&buf[i], (PAGE_SIZE - i - 2), "%u", cpu);
538 if (i >= (PAGE_SIZE - 5))
541 i += sprintf(&buf[i], "\n");
546 #define define_one_ro(_name) \
547 static struct freq_attr _name = \
548 __ATTR(_name, 0444, show_##_name, NULL)
550 #define define_one_ro0400(_name) \
551 static struct freq_attr _name = \
552 __ATTR(_name, 0400, show_##_name, NULL)
554 #define define_one_rw(_name) \
555 static struct freq_attr _name = \
556 __ATTR(_name, 0644, show_##_name, store_##_name)
558 define_one_ro0400(cpuinfo_cur_freq);
559 define_one_ro(cpuinfo_min_freq);
560 define_one_ro(cpuinfo_max_freq);
561 define_one_ro(scaling_available_governors);
562 define_one_ro(scaling_driver);
563 define_one_ro(scaling_cur_freq);
564 define_one_ro(affected_cpus);
565 define_one_rw(scaling_min_freq);
566 define_one_rw(scaling_max_freq);
567 define_one_rw(scaling_governor);
569 static struct attribute * default_attrs[] = {
570 &cpuinfo_min_freq.attr,
571 &cpuinfo_max_freq.attr,
572 &scaling_min_freq.attr,
573 &scaling_max_freq.attr,
575 &scaling_governor.attr,
576 &scaling_driver.attr,
577 &scaling_available_governors.attr,
581 #define to_policy(k) container_of(k,struct cpufreq_policy,kobj)
582 #define to_attr(a) container_of(a,struct freq_attr,attr)
584 static ssize_t show(struct kobject * kobj, struct attribute * attr ,char * buf)
586 struct cpufreq_policy * policy = to_policy(kobj);
587 struct freq_attr * fattr = to_attr(attr);
589 policy = cpufreq_cpu_get(policy->cpu);
593 ret = fattr->show(policy, buf);
597 cpufreq_cpu_put(policy);
601 static ssize_t store(struct kobject * kobj, struct attribute * attr,
602 const char * buf, size_t count)
604 struct cpufreq_policy * policy = to_policy(kobj);
605 struct freq_attr * fattr = to_attr(attr);
607 policy = cpufreq_cpu_get(policy->cpu);
611 ret = fattr->store(policy, buf, count);
615 cpufreq_cpu_put(policy);
619 static void cpufreq_sysfs_release(struct kobject * kobj)
621 struct cpufreq_policy * policy = to_policy(kobj);
622 dprintk("last reference is dropped\n");
623 complete(&policy->kobj_unregister);
626 static struct sysfs_ops sysfs_ops = {
631 static struct kobj_type ktype_cpufreq = {
632 .sysfs_ops = &sysfs_ops,
633 .default_attrs = default_attrs,
634 .release = cpufreq_sysfs_release,
639 * cpufreq_add_dev - add a CPU device
641 * Adds the cpufreq interface for a CPU device.
643 static int cpufreq_add_dev (struct sys_device * sys_dev)
645 unsigned int cpu = sys_dev->id;
647 struct cpufreq_policy new_policy;
648 struct cpufreq_policy *policy;
649 struct freq_attr **drv_attr;
650 struct sys_device *cpu_sys_dev;
654 struct cpufreq_policy *managed_policy;
657 if (cpu_is_offline(cpu))
660 cpufreq_debug_disable_ratelimit();
661 dprintk("adding CPU %u\n", cpu);
664 /* check whether a different CPU already registered this
665 * CPU because it is in the same boat. */
666 policy = cpufreq_cpu_get(cpu);
667 if (unlikely(policy)) {
668 cpufreq_cpu_put(policy);
669 cpufreq_debug_enable_ratelimit();
674 if (!try_module_get(cpufreq_driver->owner)) {
679 policy = kzalloc(sizeof(struct cpufreq_policy), GFP_KERNEL);
686 policy->cpus = cpumask_of_cpu(cpu);
688 mutex_init(&policy->lock);
689 mutex_lock(&policy->lock);
690 init_completion(&policy->kobj_unregister);
691 INIT_WORK(&policy->update, handle_update);
693 /* call driver. From then on the cpufreq must be able
694 * to accept all calls to ->verify and ->setpolicy for this CPU
696 ret = cpufreq_driver->init(policy);
698 dprintk("initialization failed\n");
699 mutex_unlock(&policy->lock);
704 for_each_cpu_mask(j, policy->cpus) {
708 /* check for existing affected CPUs. They may not be aware
709 * of it due to CPU Hotplug.
711 managed_policy = cpufreq_cpu_get(j);
712 if (unlikely(managed_policy)) {
713 spin_lock_irqsave(&cpufreq_driver_lock, flags);
714 managed_policy->cpus = policy->cpus;
715 cpufreq_cpu_data[cpu] = managed_policy;
716 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
718 dprintk("CPU already managed, adding link\n");
719 ret = sysfs_create_link(&sys_dev->kobj,
720 &managed_policy->kobj,
723 mutex_unlock(&policy->lock);
724 goto err_out_driver_exit;
727 cpufreq_debug_enable_ratelimit();
728 mutex_unlock(&policy->lock);
730 goto err_out_driver_exit; /* call driver->exit() */
734 memcpy(&new_policy, policy, sizeof(struct cpufreq_policy));
736 /* prepare interface data */
737 policy->kobj.parent = &sys_dev->kobj;
738 policy->kobj.ktype = &ktype_cpufreq;
739 strlcpy(policy->kobj.name, "cpufreq", KOBJ_NAME_LEN);
741 ret = kobject_register(&policy->kobj);
743 mutex_unlock(&policy->lock);
744 goto err_out_driver_exit;
746 /* set up files for this cpu device */
747 drv_attr = cpufreq_driver->attr;
748 while ((drv_attr) && (*drv_attr)) {
749 sysfs_create_file(&policy->kobj, &((*drv_attr)->attr));
752 if (cpufreq_driver->get)
753 sysfs_create_file(&policy->kobj, &cpuinfo_cur_freq.attr);
754 if (cpufreq_driver->target)
755 sysfs_create_file(&policy->kobj, &scaling_cur_freq.attr);
757 spin_lock_irqsave(&cpufreq_driver_lock, flags);
758 for_each_cpu_mask(j, policy->cpus)
759 cpufreq_cpu_data[j] = policy;
760 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
762 /* symlink affected CPUs */
763 for_each_cpu_mask(j, policy->cpus) {
769 dprintk("CPU %u already managed, adding link\n", j);
770 cpufreq_cpu_get(cpu);
771 cpu_sys_dev = get_cpu_sysdev(j);
772 ret = sysfs_create_link(&cpu_sys_dev->kobj, &policy->kobj,
775 mutex_unlock(&policy->lock);
776 goto err_out_unregister;
780 policy->governor = NULL; /* to assure that the starting sequence is
781 * run in cpufreq_set_policy */
782 mutex_unlock(&policy->lock);
784 /* set default policy */
785 ret = cpufreq_set_policy(&new_policy);
787 dprintk("setting policy failed\n");
788 goto err_out_unregister;
791 module_put(cpufreq_driver->owner);
792 dprintk("initialization complete\n");
793 cpufreq_debug_enable_ratelimit();
799 spin_lock_irqsave(&cpufreq_driver_lock, flags);
800 for_each_cpu_mask(j, policy->cpus)
801 cpufreq_cpu_data[j] = NULL;
802 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
804 kobject_unregister(&policy->kobj);
805 wait_for_completion(&policy->kobj_unregister);
808 if (cpufreq_driver->exit)
809 cpufreq_driver->exit(policy);
815 module_put(cpufreq_driver->owner);
817 cpufreq_debug_enable_ratelimit();
823 * cpufreq_remove_dev - remove a CPU device
825 * Removes the cpufreq interface for a CPU device.
827 static int cpufreq_remove_dev (struct sys_device * sys_dev)
829 unsigned int cpu = sys_dev->id;
831 struct cpufreq_policy *data;
833 struct sys_device *cpu_sys_dev;
837 cpufreq_debug_disable_ratelimit();
838 dprintk("unregistering CPU %u\n", cpu);
840 spin_lock_irqsave(&cpufreq_driver_lock, flags);
841 data = cpufreq_cpu_data[cpu];
844 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
845 cpufreq_debug_enable_ratelimit();
848 cpufreq_cpu_data[cpu] = NULL;
852 /* if this isn't the CPU which is the parent of the kobj, we
853 * only need to unlink, put and exit
855 if (unlikely(cpu != data->cpu)) {
856 dprintk("removing link\n");
857 cpu_clear(cpu, data->cpus);
858 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
859 sysfs_remove_link(&sys_dev->kobj, "cpufreq");
860 cpufreq_cpu_put(data);
861 cpufreq_debug_enable_ratelimit();
867 if (!kobject_get(&data->kobj)) {
868 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
869 cpufreq_debug_enable_ratelimit();
874 /* if we have other CPUs still registered, we need to unlink them,
875 * or else wait_for_completion below will lock up. Clean the
876 * cpufreq_cpu_data[] while holding the lock, and remove the sysfs
879 if (unlikely(cpus_weight(data->cpus) > 1)) {
880 for_each_cpu_mask(j, data->cpus) {
883 cpufreq_cpu_data[j] = NULL;
887 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
889 if (unlikely(cpus_weight(data->cpus) > 1)) {
890 for_each_cpu_mask(j, data->cpus) {
893 dprintk("removing link for cpu %u\n", j);
894 cpu_sys_dev = get_cpu_sysdev(j);
895 sysfs_remove_link(&cpu_sys_dev->kobj, "cpufreq");
896 cpufreq_cpu_put(data);
900 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
903 mutex_lock(&data->lock);
904 if (cpufreq_driver->target)
905 __cpufreq_governor(data, CPUFREQ_GOV_STOP);
906 mutex_unlock(&data->lock);
908 kobject_unregister(&data->kobj);
910 kobject_put(&data->kobj);
912 /* we need to make sure that the underlying kobj is actually
913 * not referenced anymore by anybody before we proceed with
916 dprintk("waiting for dropping of refcount\n");
917 wait_for_completion(&data->kobj_unregister);
918 dprintk("wait complete\n");
920 if (cpufreq_driver->exit)
921 cpufreq_driver->exit(data);
925 cpufreq_debug_enable_ratelimit();
930 static void handle_update(struct work_struct *work)
932 struct cpufreq_policy *policy =
933 container_of(work, struct cpufreq_policy, update);
934 unsigned int cpu = policy->cpu;
935 dprintk("handle_update for cpu %u called\n", cpu);
936 cpufreq_update_policy(cpu);
940 * cpufreq_out_of_sync - If actual and saved CPU frequency differs, we're in deep trouble.
942 * @old_freq: CPU frequency the kernel thinks the CPU runs at
943 * @new_freq: CPU frequency the CPU actually runs at
945 * We adjust to current frequency first, and need to clean up later. So either call
946 * to cpufreq_update_policy() or schedule handle_update()).
948 static void cpufreq_out_of_sync(unsigned int cpu, unsigned int old_freq,
949 unsigned int new_freq)
951 struct cpufreq_freqs freqs;
953 dprintk("Warning: CPU frequency out of sync: cpufreq and timing "
954 "core thinks of %u, is %u kHz.\n", old_freq, new_freq);
957 freqs.old = old_freq;
958 freqs.new = new_freq;
959 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
960 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
965 * cpufreq_quick_get - get the CPU frequency (in kHz) from policy->cur
968 * This is the last known freq, without actually getting it from the driver.
969 * Return value will be same as what is shown in scaling_cur_freq in sysfs.
971 unsigned int cpufreq_quick_get(unsigned int cpu)
973 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
974 unsigned int ret_freq = 0;
977 mutex_lock(&policy->lock);
978 ret_freq = policy->cur;
979 mutex_unlock(&policy->lock);
980 cpufreq_cpu_put(policy);
985 EXPORT_SYMBOL(cpufreq_quick_get);
989 * cpufreq_get - get the current CPU frequency (in kHz)
992 * Get the CPU current (static) CPU frequency
994 unsigned int cpufreq_get(unsigned int cpu)
996 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
997 unsigned int ret_freq = 0;
1002 if (!cpufreq_driver->get)
1005 mutex_lock(&policy->lock);
1007 ret_freq = cpufreq_driver->get(cpu);
1009 if (ret_freq && policy->cur &&
1010 !(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
1011 /* verify no discrepancy between actual and
1012 saved value exists */
1013 if (unlikely(ret_freq != policy->cur)) {
1014 cpufreq_out_of_sync(cpu, policy->cur, ret_freq);
1015 schedule_work(&policy->update);
1019 mutex_unlock(&policy->lock);
1022 cpufreq_cpu_put(policy);
1026 EXPORT_SYMBOL(cpufreq_get);
1030 * cpufreq_suspend - let the low level driver prepare for suspend
1033 static int cpufreq_suspend(struct sys_device * sysdev, pm_message_t pmsg)
1035 int cpu = sysdev->id;
1037 unsigned int cur_freq = 0;
1038 struct cpufreq_policy *cpu_policy;
1040 dprintk("suspending cpu %u\n", cpu);
1042 if (!cpu_online(cpu))
1045 /* we may be lax here as interrupts are off. Nonetheless
1046 * we need to grab the correct cpu policy, as to check
1047 * whether we really run on this CPU.
1050 cpu_policy = cpufreq_cpu_get(cpu);
1054 /* only handle each CPU group once */
1055 if (unlikely(cpu_policy->cpu != cpu)) {
1056 cpufreq_cpu_put(cpu_policy);
1060 if (cpufreq_driver->suspend) {
1061 ret = cpufreq_driver->suspend(cpu_policy, pmsg);
1063 printk(KERN_ERR "cpufreq: suspend failed in ->suspend "
1064 "step on CPU %u\n", cpu_policy->cpu);
1065 cpufreq_cpu_put(cpu_policy);
1071 if (cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)
1074 if (cpufreq_driver->get)
1075 cur_freq = cpufreq_driver->get(cpu_policy->cpu);
1077 if (!cur_freq || !cpu_policy->cur) {
1078 printk(KERN_ERR "cpufreq: suspend failed to assert current "
1079 "frequency is what timing core thinks it is.\n");
1083 if (unlikely(cur_freq != cpu_policy->cur)) {
1084 struct cpufreq_freqs freqs;
1086 if (!(cpufreq_driver->flags & CPUFREQ_PM_NO_WARN))
1087 dprintk("Warning: CPU frequency is %u, "
1088 "cpufreq assumed %u kHz.\n",
1089 cur_freq, cpu_policy->cur);
1092 freqs.old = cpu_policy->cur;
1093 freqs.new = cur_freq;
1095 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
1096 CPUFREQ_SUSPENDCHANGE, &freqs);
1097 adjust_jiffies(CPUFREQ_SUSPENDCHANGE, &freqs);
1099 cpu_policy->cur = cur_freq;
1103 cpufreq_cpu_put(cpu_policy);
1108 * cpufreq_resume - restore proper CPU frequency handling after resume
1110 * 1.) resume CPUfreq hardware support (cpufreq_driver->resume())
1111 * 2.) if ->target and !CPUFREQ_CONST_LOOPS: verify we're in sync
1112 * 3.) schedule call cpufreq_update_policy() ASAP as interrupts are
1115 static int cpufreq_resume(struct sys_device * sysdev)
1117 int cpu = sysdev->id;
1119 struct cpufreq_policy *cpu_policy;
1121 dprintk("resuming cpu %u\n", cpu);
1123 if (!cpu_online(cpu))
1126 /* we may be lax here as interrupts are off. Nonetheless
1127 * we need to grab the correct cpu policy, as to check
1128 * whether we really run on this CPU.
1131 cpu_policy = cpufreq_cpu_get(cpu);
1135 /* only handle each CPU group once */
1136 if (unlikely(cpu_policy->cpu != cpu)) {
1137 cpufreq_cpu_put(cpu_policy);
1141 if (cpufreq_driver->resume) {
1142 ret = cpufreq_driver->resume(cpu_policy);
1144 printk(KERN_ERR "cpufreq: resume failed in ->resume "
1145 "step on CPU %u\n", cpu_policy->cpu);
1146 cpufreq_cpu_put(cpu_policy);
1151 if (!(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
1152 unsigned int cur_freq = 0;
1154 if (cpufreq_driver->get)
1155 cur_freq = cpufreq_driver->get(cpu_policy->cpu);
1157 if (!cur_freq || !cpu_policy->cur) {
1158 printk(KERN_ERR "cpufreq: resume failed to assert "
1159 "current frequency is what timing core "
1164 if (unlikely(cur_freq != cpu_policy->cur)) {
1165 struct cpufreq_freqs freqs;
1167 if (!(cpufreq_driver->flags & CPUFREQ_PM_NO_WARN))
1168 dprintk("Warning: CPU frequency"
1169 "is %u, cpufreq assumed %u kHz.\n",
1170 cur_freq, cpu_policy->cur);
1173 freqs.old = cpu_policy->cur;
1174 freqs.new = cur_freq;
1176 srcu_notifier_call_chain(
1177 &cpufreq_transition_notifier_list,
1178 CPUFREQ_RESUMECHANGE, &freqs);
1179 adjust_jiffies(CPUFREQ_RESUMECHANGE, &freqs);
1181 cpu_policy->cur = cur_freq;
1186 schedule_work(&cpu_policy->update);
1187 cpufreq_cpu_put(cpu_policy);
1191 static struct sysdev_driver cpufreq_sysdev_driver = {
1192 .add = cpufreq_add_dev,
1193 .remove = cpufreq_remove_dev,
1194 .suspend = cpufreq_suspend,
1195 .resume = cpufreq_resume,
1199 /*********************************************************************
1200 * NOTIFIER LISTS INTERFACE *
1201 *********************************************************************/
1204 * cpufreq_register_notifier - register a driver with cpufreq
1205 * @nb: notifier function to register
1206 * @list: CPUFREQ_TRANSITION_NOTIFIER or CPUFREQ_POLICY_NOTIFIER
1208 * Add a driver to one of two lists: either a list of drivers that
1209 * are notified about clock rate changes (once before and once after
1210 * the transition), or a list of drivers that are notified about
1211 * changes in cpufreq policy.
1213 * This function may sleep, and has the same return conditions as
1214 * blocking_notifier_chain_register.
1216 int cpufreq_register_notifier(struct notifier_block *nb, unsigned int list)
1221 case CPUFREQ_TRANSITION_NOTIFIER:
1222 ret = srcu_notifier_chain_register(
1223 &cpufreq_transition_notifier_list, nb);
1225 case CPUFREQ_POLICY_NOTIFIER:
1226 ret = blocking_notifier_chain_register(
1227 &cpufreq_policy_notifier_list, nb);
1235 EXPORT_SYMBOL(cpufreq_register_notifier);
1239 * cpufreq_unregister_notifier - unregister a driver with cpufreq
1240 * @nb: notifier block to be unregistered
1241 * @list: CPUFREQ_TRANSITION_NOTIFIER or CPUFREQ_POLICY_NOTIFIER
1243 * Remove a driver from the CPU frequency notifier list.
1245 * This function may sleep, and has the same return conditions as
1246 * blocking_notifier_chain_unregister.
1248 int cpufreq_unregister_notifier(struct notifier_block *nb, unsigned int list)
1253 case CPUFREQ_TRANSITION_NOTIFIER:
1254 ret = srcu_notifier_chain_unregister(
1255 &cpufreq_transition_notifier_list, nb);
1257 case CPUFREQ_POLICY_NOTIFIER:
1258 ret = blocking_notifier_chain_unregister(
1259 &cpufreq_policy_notifier_list, nb);
1267 EXPORT_SYMBOL(cpufreq_unregister_notifier);
1270 /*********************************************************************
1272 *********************************************************************/
1275 int __cpufreq_driver_target(struct cpufreq_policy *policy,
1276 unsigned int target_freq,
1277 unsigned int relation)
1279 int retval = -EINVAL;
1281 dprintk("target for CPU %u: %u kHz, relation %u\n", policy->cpu,
1282 target_freq, relation);
1283 if (cpu_online(policy->cpu) && cpufreq_driver->target)
1284 retval = cpufreq_driver->target(policy, target_freq, relation);
1288 EXPORT_SYMBOL_GPL(__cpufreq_driver_target);
1290 int cpufreq_driver_target(struct cpufreq_policy *policy,
1291 unsigned int target_freq,
1292 unsigned int relation)
1296 policy = cpufreq_cpu_get(policy->cpu);
1300 mutex_lock(&policy->lock);
1302 ret = __cpufreq_driver_target(policy, target_freq, relation);
1304 mutex_unlock(&policy->lock);
1306 cpufreq_cpu_put(policy);
1309 EXPORT_SYMBOL_GPL(cpufreq_driver_target);
1311 int cpufreq_driver_getavg(struct cpufreq_policy *policy)
1315 policy = cpufreq_cpu_get(policy->cpu);
1319 mutex_lock(&policy->lock);
1321 if (cpu_online(policy->cpu) && cpufreq_driver->getavg)
1322 ret = cpufreq_driver->getavg(policy->cpu);
1324 mutex_unlock(&policy->lock);
1326 cpufreq_cpu_put(policy);
1329 EXPORT_SYMBOL_GPL(cpufreq_driver_getavg);
1332 * when "event" is CPUFREQ_GOV_LIMITS
1335 static int __cpufreq_governor(struct cpufreq_policy *policy,
1340 if (!try_module_get(policy->governor->owner))
1343 dprintk("__cpufreq_governor for CPU %u, event %u\n",
1344 policy->cpu, event);
1345 ret = policy->governor->governor(policy, event);
1347 /* we keep one module reference alive for
1348 each CPU governed by this CPU */
1349 if ((event != CPUFREQ_GOV_START) || ret)
1350 module_put(policy->governor->owner);
1351 if ((event == CPUFREQ_GOV_STOP) && !ret)
1352 module_put(policy->governor->owner);
1358 int cpufreq_register_governor(struct cpufreq_governor *governor)
1365 mutex_lock(&cpufreq_governor_mutex);
1368 if (__find_governor(governor->name) == NULL) {
1370 list_add(&governor->governor_list, &cpufreq_governor_list);
1373 mutex_unlock(&cpufreq_governor_mutex);
1376 EXPORT_SYMBOL_GPL(cpufreq_register_governor);
1379 void cpufreq_unregister_governor(struct cpufreq_governor *governor)
1384 mutex_lock(&cpufreq_governor_mutex);
1385 list_del(&governor->governor_list);
1386 mutex_unlock(&cpufreq_governor_mutex);
1389 EXPORT_SYMBOL_GPL(cpufreq_unregister_governor);
1393 /*********************************************************************
1394 * POLICY INTERFACE *
1395 *********************************************************************/
1398 * cpufreq_get_policy - get the current cpufreq_policy
1399 * @policy: struct cpufreq_policy into which the current cpufreq_policy is written
1401 * Reads the current cpufreq policy.
1403 int cpufreq_get_policy(struct cpufreq_policy *policy, unsigned int cpu)
1405 struct cpufreq_policy *cpu_policy;
1409 cpu_policy = cpufreq_cpu_get(cpu);
1413 mutex_lock(&cpu_policy->lock);
1414 memcpy(policy, cpu_policy, sizeof(struct cpufreq_policy));
1415 mutex_unlock(&cpu_policy->lock);
1417 cpufreq_cpu_put(cpu_policy);
1420 EXPORT_SYMBOL(cpufreq_get_policy);
1424 * data : current policy.
1425 * policy : policy to be set.
1427 static int __cpufreq_set_policy(struct cpufreq_policy *data,
1428 struct cpufreq_policy *policy)
1432 cpufreq_debug_disable_ratelimit();
1433 dprintk("setting new policy for CPU %u: %u - %u kHz\n", policy->cpu,
1434 policy->min, policy->max);
1436 memcpy(&policy->cpuinfo, &data->cpuinfo,
1437 sizeof(struct cpufreq_cpuinfo));
1439 if (policy->min > data->min && policy->min > policy->max) {
1444 /* verify the cpu speed can be set within this limit */
1445 ret = cpufreq_driver->verify(policy);
1449 /* adjust if necessary - all reasons */
1450 blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1451 CPUFREQ_ADJUST, policy);
1453 /* adjust if necessary - hardware incompatibility*/
1454 blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1455 CPUFREQ_INCOMPATIBLE, policy);
1457 /* verify the cpu speed can be set within this limit,
1458 which might be different to the first one */
1459 ret = cpufreq_driver->verify(policy);
1463 /* notification of the new policy */
1464 blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1465 CPUFREQ_NOTIFY, policy);
1467 data->min = policy->min;
1468 data->max = policy->max;
1470 dprintk("new min and max freqs are %u - %u kHz\n",
1471 data->min, data->max);
1473 if (cpufreq_driver->setpolicy) {
1474 data->policy = policy->policy;
1475 dprintk("setting range\n");
1476 ret = cpufreq_driver->setpolicy(policy);
1478 if (policy->governor != data->governor) {
1479 /* save old, working values */
1480 struct cpufreq_governor *old_gov = data->governor;
1482 dprintk("governor switch\n");
1484 /* end old governor */
1486 __cpufreq_governor(data, CPUFREQ_GOV_STOP);
1488 /* start new governor */
1489 data->governor = policy->governor;
1490 if (__cpufreq_governor(data, CPUFREQ_GOV_START)) {
1491 /* new governor failed, so re-start old one */
1492 dprintk("starting governor %s failed\n",
1493 data->governor->name);
1495 data->governor = old_gov;
1496 __cpufreq_governor(data,
1502 /* might be a policy change, too, so fall through */
1504 dprintk("governor: change or update limits\n");
1505 __cpufreq_governor(data, CPUFREQ_GOV_LIMITS);
1509 cpufreq_debug_enable_ratelimit();
1514 * cpufreq_set_policy - set a new CPUFreq policy
1515 * @policy: policy to be set.
1517 * Sets a new CPU frequency and voltage scaling policy.
1519 int cpufreq_set_policy(struct cpufreq_policy *policy)
1522 struct cpufreq_policy *data;
1527 data = cpufreq_cpu_get(policy->cpu);
1532 mutex_lock(&data->lock);
1534 ret = __cpufreq_set_policy(data, policy);
1535 data->user_policy.min = data->min;
1536 data->user_policy.max = data->max;
1537 data->user_policy.policy = data->policy;
1538 data->user_policy.governor = data->governor;
1540 mutex_unlock(&data->lock);
1542 cpufreq_cpu_put(data);
1546 EXPORT_SYMBOL(cpufreq_set_policy);
1550 * cpufreq_update_policy - re-evaluate an existing cpufreq policy
1551 * @cpu: CPU which shall be re-evaluated
1553 * Usefull for policy notifiers which have different necessities
1554 * at different times.
1556 int cpufreq_update_policy(unsigned int cpu)
1558 struct cpufreq_policy *data = cpufreq_cpu_get(cpu);
1559 struct cpufreq_policy policy;
1565 mutex_lock(&data->lock);
1567 dprintk("updating policy for CPU %u\n", cpu);
1568 memcpy(&policy, data, sizeof(struct cpufreq_policy));
1569 policy.min = data->user_policy.min;
1570 policy.max = data->user_policy.max;
1571 policy.policy = data->user_policy.policy;
1572 policy.governor = data->user_policy.governor;
1574 /* BIOS might change freq behind our back
1575 -> ask driver for current freq and notify governors about a change */
1576 if (cpufreq_driver->get) {
1577 policy.cur = cpufreq_driver->get(cpu);
1579 dprintk("Driver did not initialize current freq");
1580 data->cur = policy.cur;
1582 if (data->cur != policy.cur)
1583 cpufreq_out_of_sync(cpu, data->cur,
1588 ret = __cpufreq_set_policy(data, &policy);
1590 mutex_unlock(&data->lock);
1591 cpufreq_cpu_put(data);
1594 EXPORT_SYMBOL(cpufreq_update_policy);
1596 static int cpufreq_cpu_callback(struct notifier_block *nfb,
1597 unsigned long action, void *hcpu)
1599 unsigned int cpu = (unsigned long)hcpu;
1600 struct cpufreq_policy *policy;
1601 struct sys_device *sys_dev;
1603 sys_dev = get_cpu_sysdev(cpu);
1608 cpufreq_add_dev(sys_dev);
1610 case CPU_DOWN_PREPARE:
1612 * We attempt to put this cpu in lowest frequency
1613 * possible before going down. This will permit
1614 * hardware-managed P-State to switch other related
1615 * threads to min or higher speeds if possible.
1617 policy = cpufreq_cpu_data[cpu];
1619 cpufreq_driver_target(policy, policy->min,
1620 CPUFREQ_RELATION_H);
1624 cpufreq_remove_dev(sys_dev);
1631 static struct notifier_block __cpuinitdata cpufreq_cpu_notifier =
1633 .notifier_call = cpufreq_cpu_callback,
1636 /*********************************************************************
1637 * REGISTER / UNREGISTER CPUFREQ DRIVER *
1638 *********************************************************************/
1641 * cpufreq_register_driver - register a CPU Frequency driver
1642 * @driver_data: A struct cpufreq_driver containing the values#
1643 * submitted by the CPU Frequency driver.
1645 * Registers a CPU Frequency driver to this core code. This code
1646 * returns zero on success, -EBUSY when another driver got here first
1647 * (and isn't unregistered in the meantime).
1650 int cpufreq_register_driver(struct cpufreq_driver *driver_data)
1652 unsigned long flags;
1655 if (!driver_data || !driver_data->verify || !driver_data->init ||
1656 ((!driver_data->setpolicy) && (!driver_data->target)))
1659 dprintk("trying to register driver %s\n", driver_data->name);
1661 if (driver_data->setpolicy)
1662 driver_data->flags |= CPUFREQ_CONST_LOOPS;
1664 spin_lock_irqsave(&cpufreq_driver_lock, flags);
1665 if (cpufreq_driver) {
1666 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1669 cpufreq_driver = driver_data;
1670 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1672 ret = sysdev_driver_register(&cpu_sysdev_class,&cpufreq_sysdev_driver);
1674 if ((!ret) && !(cpufreq_driver->flags & CPUFREQ_STICKY)) {
1678 /* check for at least one working CPU */
1679 for (i=0; i<NR_CPUS; i++)
1680 if (cpufreq_cpu_data[i])
1683 /* if all ->init() calls failed, unregister */
1685 dprintk("no CPU initialized for driver %s\n",
1687 sysdev_driver_unregister(&cpu_sysdev_class,
1688 &cpufreq_sysdev_driver);
1690 spin_lock_irqsave(&cpufreq_driver_lock, flags);
1691 cpufreq_driver = NULL;
1692 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1697 register_hotcpu_notifier(&cpufreq_cpu_notifier);
1698 dprintk("driver %s up and running\n", driver_data->name);
1699 cpufreq_debug_enable_ratelimit();
1704 EXPORT_SYMBOL_GPL(cpufreq_register_driver);
1708 * cpufreq_unregister_driver - unregister the current CPUFreq driver
1710 * Unregister the current CPUFreq driver. Only call this if you have
1711 * the right to do so, i.e. if you have succeeded in initialising before!
1712 * Returns zero if successful, and -EINVAL if the cpufreq_driver is
1713 * currently not initialised.
1715 int cpufreq_unregister_driver(struct cpufreq_driver *driver)
1717 unsigned long flags;
1719 cpufreq_debug_disable_ratelimit();
1721 if (!cpufreq_driver || (driver != cpufreq_driver)) {
1722 cpufreq_debug_enable_ratelimit();
1726 dprintk("unregistering driver %s\n", driver->name);
1728 sysdev_driver_unregister(&cpu_sysdev_class, &cpufreq_sysdev_driver);
1729 unregister_hotcpu_notifier(&cpufreq_cpu_notifier);
1731 spin_lock_irqsave(&cpufreq_driver_lock, flags);
1732 cpufreq_driver = NULL;
1733 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1737 EXPORT_SYMBOL_GPL(cpufreq_unregister_driver);