git.oblomov.eu Git - linux-2.6/blob - 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(struct cpufreq_governor *, 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  */
  65 static DEFINE_PER_CPU(int, policy_cpu);
  66 static DEFINE_PER_CPU(struct rw_semaphore, cpu_policy_rwsem);
  67
  68 #define lock_policy_rwsem(mode, cpu)                                    \
  69 int lock_policy_rwsem_##mode                                            \
  70 (int cpu)                                                               \
  71 {                                                                       \
  72         int policy_cpu = per_cpu(policy_cpu, cpu);                      \
  73         BUG_ON(policy_cpu == -1);                                       \
  74         down_##mode(&per_cpu(cpu_policy_rwsem, policy_cpu));            \
  75         if (unlikely(!cpu_online(cpu))) {                               \
  76                 up_##mode(&per_cpu(cpu_policy_rwsem, policy_cpu));      \
  77                 return -1;                                              \
  78         }                                                               \
  79                                                                         \
  80         return 0;                                                       \
  81 }
  82
  83 lock_policy_rwsem(read, cpu);
  84 EXPORT_SYMBOL_GPL(lock_policy_rwsem_read);
  85
  86 lock_policy_rwsem(write, cpu);
  87 EXPORT_SYMBOL_GPL(lock_policy_rwsem_write);
  88
  89 void unlock_policy_rwsem_read(int cpu)
  90 {
  91         int policy_cpu = per_cpu(policy_cpu, cpu);
  92         BUG_ON(policy_cpu == -1);
  93         up_read(&per_cpu(cpu_policy_rwsem, policy_cpu));
  94 }
  95 EXPORT_SYMBOL_GPL(unlock_policy_rwsem_read);
  96
  97 void unlock_policy_rwsem_write(int cpu)
  98 {
  99         int policy_cpu = per_cpu(policy_cpu, cpu);
 100         BUG_ON(policy_cpu == -1);
 101         up_write(&per_cpu(cpu_policy_rwsem, policy_cpu));
 102 }
 103 EXPORT_SYMBOL_GPL(unlock_policy_rwsem_write);
 104
 105
 106 /* internal prototypes */
 107 static int __cpufreq_governor(struct cpufreq_policy *policy, unsigned int event);
 108 static unsigned int __cpufreq_get(unsigned int cpu);
 109 static void handle_update(struct work_struct *work);
 110
 111 /**
 112  * Two notifier lists: the "policy" list is involved in the
 113  * validation process for a new CPU frequency policy; the
 114  * "transition" list for kernel code that needs to handle
 115  * changes to devices when the CPU clock speed changes.
 116  * The mutex locks both lists.
 117  */
 118 static BLOCKING_NOTIFIER_HEAD(cpufreq_policy_notifier_list);
 119 static struct srcu_notifier_head cpufreq_transition_notifier_list;
 120
 121 static bool init_cpufreq_transition_notifier_list_called;
 122 static int __init init_cpufreq_transition_notifier_list(void)
 123 {
 124         srcu_init_notifier_head(&cpufreq_transition_notifier_list);
 125         init_cpufreq_transition_notifier_list_called = true;
 126         return 0;
 127 }
 128 pure_initcall(init_cpufreq_transition_notifier_list);
 129
 130 static LIST_HEAD(cpufreq_governor_list);
 131 static DEFINE_MUTEX (cpufreq_governor_mutex);
 132
 133 struct cpufreq_policy *cpufreq_cpu_get(unsigned int cpu)
 134 {
 135         struct cpufreq_policy *data;
 136         unsigned long flags;
 137
 138         if (cpu >= nr_cpu_ids)
 139                 goto err_out;
 140
 141         /* get the cpufreq driver */
 142         spin_lock_irqsave(&cpufreq_driver_lock, flags);
 143
 144         if (!cpufreq_driver)
 145                 goto err_out_unlock;
 146
 147         if (!try_module_get(cpufreq_driver->owner))
 148                 goto err_out_unlock;
 149
 150
 151         /* get the CPU */
 152         data = per_cpu(cpufreq_cpu_data, cpu);
 153
 154         if (!data)
 155                 goto err_out_put_module;
 156
 157         if (!kobject_get(&data->kobj))
 158                 goto err_out_put_module;
 159
 160         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
 161         return data;
 162
 163 err_out_put_module:
 164         module_put(cpufreq_driver->owner);
 165 err_out_unlock:
 166         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
 167 err_out:
 168         return NULL;
 169 }
 170 EXPORT_SYMBOL_GPL(cpufreq_cpu_get);
 171
 172
 173 void cpufreq_cpu_put(struct cpufreq_policy *data)
 174 {
 175         kobject_put(&data->kobj);
 176         module_put(cpufreq_driver->owner);
 177 }
 178 EXPORT_SYMBOL_GPL(cpufreq_cpu_put);
 179
 180
 181 /*********************************************************************
 182  *                     UNIFIED DEBUG HELPERS                         *
 183  *********************************************************************/
 184 #ifdef CONFIG_CPU_FREQ_DEBUG
 185
 186 /* what part(s) of the CPUfreq subsystem are debugged? */
 187 static unsigned int debug;
 188
 189 /* is the debug output ratelimit'ed using printk_ratelimit? User can
 190  * set or modify this value.
 191  */
 192 static unsigned int debug_ratelimit = 1;
 193
 194 /* is the printk_ratelimit'ing enabled? It's enabled after a successful
 195  * loading of a cpufreq driver, temporarily disabled when a new policy
 196  * is set, and disabled upon cpufreq driver removal
 197  */
 198 static unsigned int disable_ratelimit = 1;
 199 static DEFINE_SPINLOCK(disable_ratelimit_lock);
 200
 201 static void cpufreq_debug_enable_ratelimit(void)
 202 {
 203         unsigned long flags;
 204
 205         spin_lock_irqsave(&disable_ratelimit_lock, flags);
 206         if (disable_ratelimit)
 207                 disable_ratelimit--;
 208         spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
 209 }
 210
 211 static void cpufreq_debug_disable_ratelimit(void)
 212 {
 213         unsigned long flags;
 214
 215         spin_lock_irqsave(&disable_ratelimit_lock, flags);
 216         disable_ratelimit++;
 217         spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
 218 }
 219
 220 void cpufreq_debug_printk(unsigned int type, const char *prefix,
 221                         const char *fmt, ...)
 222 {
 223         char s[256];
 224         va_list args;
 225         unsigned int len;
 226         unsigned long flags;
 227
 228         WARN_ON(!prefix);
 229         if (type & debug) {
 230                 spin_lock_irqsave(&disable_ratelimit_lock, flags);
 231                 if (!disable_ratelimit && debug_ratelimit
 232                                         && !printk_ratelimit()) {
 233                         spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
 234                         return;
 235                 }
 236                 spin_unlock_irqrestore(&disable_ratelimit_lock, flags);
 237
 238                 len = snprintf(s, 256, KERN_DEBUG "%s: ", prefix);
 239
 240                 va_start(args, fmt);
 241                 len += vsnprintf(&s[len], (256 - len), fmt, args);
 242                 va_end(args);
 243
 244                 printk(s);
 245
 246                 WARN_ON(len < 5);
 247         }
 248 }
 249 EXPORT_SYMBOL(cpufreq_debug_printk);
 250
 251
 252 module_param(debug, uint, 0644);
 253 MODULE_PARM_DESC(debug, "CPUfreq debugging: add 1 to debug core,"
 254                         " 2 to debug drivers, and 4 to debug governors.");
 255
 256 module_param(debug_ratelimit, uint, 0644);
 257 MODULE_PARM_DESC(debug_ratelimit, "CPUfreq debugging:"
 258                                         " set to 0 to disable ratelimiting.");
 259
 260 #else /* !CONFIG_CPU_FREQ_DEBUG */
 261
 262 static inline void cpufreq_debug_enable_ratelimit(void) { return; }
 263 static inline void cpufreq_debug_disable_ratelimit(void) { return; }
 264
 265 #endif /* CONFIG_CPU_FREQ_DEBUG */
 266
 267
 268 /*********************************************************************
 269  *            EXTERNALLY AFFECTING FREQUENCY CHANGES                 *
 270  *********************************************************************/
 271
 272 /**
 273  * adjust_jiffies - adjust the system "loops_per_jiffy"
 274  *
 275  * This function alters the system "loops_per_jiffy" for the clock
 276  * speed change. Note that loops_per_jiffy cannot be updated on SMP
 277  * systems as each CPU might be scaled differently. So, use the arch
 278  * per-CPU loops_per_jiffy value wherever possible.
 279  */
 280 #ifndef CONFIG_SMP
 281 static unsigned long l_p_j_ref;
 282 static unsigned int  l_p_j_ref_freq;
 283
 284 static void adjust_jiffies(unsigned long val, struct cpufreq_freqs *ci)
 285 {
 286         if (ci->flags & CPUFREQ_CONST_LOOPS)
 287                 return;
 288
 289         if (!l_p_j_ref_freq) {
 290                 l_p_j_ref = loops_per_jiffy;
 291                 l_p_j_ref_freq = ci->old;
 292                 dprintk("saving %lu as reference value for loops_per_jiffy; "
 293                         "freq is %u kHz\n", l_p_j_ref, l_p_j_ref_freq);
 294         }
 295         if ((val == CPUFREQ_PRECHANGE  && ci->old < ci->new) ||
 296             (val == CPUFREQ_POSTCHANGE && ci->old > ci->new) ||
 297             (val == CPUFREQ_RESUMECHANGE || val == CPUFREQ_SUSPENDCHANGE)) {
 298                 loops_per_jiffy = cpufreq_scale(l_p_j_ref, l_p_j_ref_freq,
 299                                                                 ci->new);
 300                 dprintk("scaling loops_per_jiffy to %lu "
 301                         "for frequency %u kHz\n", loops_per_jiffy, ci->new);
 302         }
 303 }
 304 #else
 305 static inline void adjust_jiffies(unsigned long val, struct cpufreq_freqs *ci)
 306 {
 307         return;
 308 }
 309 #endif
 310
 311
 312 /**
 313  * cpufreq_notify_transition - call notifier chain and adjust_jiffies
 314  * on frequency transition.
 315  *
 316  * This function calls the transition notifiers and the "adjust_jiffies"
 317  * function. It is called twice on all CPU frequency changes that have
 318  * external effects.
 319  */
 320 void cpufreq_notify_transition(struct cpufreq_freqs *freqs, unsigned int state)
 321 {
 322         struct cpufreq_policy *policy;
 323
 324         BUG_ON(irqs_disabled());
 325
 326         freqs->flags = cpufreq_driver->flags;
 327         dprintk("notification %u of frequency transition to %u kHz\n",
 328                 state, freqs->new);
 329
 330         policy = per_cpu(cpufreq_cpu_data, freqs->cpu);
 331         switch (state) {
 332
 333         case CPUFREQ_PRECHANGE:
 334                 /* detect if the driver reported a value as "old frequency"
 335                  * which is not equal to what the cpufreq core thinks is
 336                  * "old frequency".
 337                  */
 338                 if (!(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
 339                         if ((policy) && (policy->cpu == freqs->cpu) &&
 340                             (policy->cur) && (policy->cur != freqs->old)) {
 341                                 dprintk("Warning: CPU frequency is"
 342                                         " %u, cpufreq assumed %u kHz.\n",
 343                                         freqs->old, policy->cur);
 344                                 freqs->old = policy->cur;
 345                         }
 346                 }
 347                 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
 348                                 CPUFREQ_PRECHANGE, freqs);
 349                 adjust_jiffies(CPUFREQ_PRECHANGE, freqs);
 350                 break;
 351
 352         case CPUFREQ_POSTCHANGE:
 353                 adjust_jiffies(CPUFREQ_POSTCHANGE, freqs);
 354                 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
 355                                 CPUFREQ_POSTCHANGE, freqs);
 356                 if (likely(policy) && likely(policy->cpu == freqs->cpu))
 357                         policy->cur = freqs->new;
 358                 break;
 359         }
 360 }
 361 EXPORT_SYMBOL_GPL(cpufreq_notify_transition);
 362
 363
 364
 365 /*********************************************************************
 366  *                          SYSFS INTERFACE                          *
 367  *********************************************************************/
 368
 369 static struct cpufreq_governor *__find_governor(const char *str_governor)
 370 {
 371         struct cpufreq_governor *t;
 372
 373         list_for_each_entry(t, &cpufreq_governor_list, governor_list)
 374                 if (!strnicmp(str_governor,t->name,CPUFREQ_NAME_LEN))
 375                         return t;
 376
 377         return NULL;
 378 }
 379
 380 /**
 381  * cpufreq_parse_governor - parse a governor string
 382  */
 383 static int cpufreq_parse_governor(char *str_governor, unsigned int *policy,
 384                                 struct cpufreq_governor **governor)
 385 {
 386         int err = -EINVAL;
 387
 388         if (!cpufreq_driver)
 389                 goto out;
 390
 391         if (cpufreq_driver->setpolicy) {
 392                 if (!strnicmp(str_governor, "performance", CPUFREQ_NAME_LEN)) {
 393                         *policy = CPUFREQ_POLICY_PERFORMANCE;
 394                         err = 0;
 395                 } else if (!strnicmp(str_governor, "powersave",
 396                                                 CPUFREQ_NAME_LEN)) {
 397                         *policy = CPUFREQ_POLICY_POWERSAVE;
 398                         err = 0;
 399                 }
 400         } else if (cpufreq_driver->target) {
 401                 struct cpufreq_governor *t;
 402
 403                 mutex_lock(&cpufreq_governor_mutex);
 404
 405                 t = __find_governor(str_governor);
 406
 407                 if (t == NULL) {
 408                         char *name = kasprintf(GFP_KERNEL, "cpufreq_%s",
 409                                                                 str_governor);
 410
 411                         if (name) {
 412                                 int ret;
 413
 414                                 mutex_unlock(&cpufreq_governor_mutex);
 415                                 ret = request_module("%s", name);
 416                                 mutex_lock(&cpufreq_governor_mutex);
 417
 418                                 if (ret == 0)
 419                                         t = __find_governor(str_governor);
 420                         }
 421
 422                         kfree(name);
 423                 }
 424
 425                 if (t != NULL) {
 426                         *governor = t;
 427                         err = 0;
 428                 }
 429
 430                 mutex_unlock(&cpufreq_governor_mutex);
 431         }
 432   out:
 433         return err;
 434 }
 435
 436
 437 /* drivers/base/cpu.c */
 438 extern struct sysdev_class cpu_sysdev_class;
 439
 440
 441 /**
 442  * cpufreq_per_cpu_attr_read() / show_##file_name() -
 443  * print out cpufreq information
 444  *
 445  * Write out information from cpufreq_driver->policy[cpu]; object must be
 446  * "unsigned int".
 447  */
 448
 449 #define show_one(file_name, object)                     \
 450 static ssize_t show_##file_name                         \
 451 (struct cpufreq_policy *policy, char *buf)              \
 452 {                                                       \
 453         return sprintf (buf, "%u\n", policy->object);   \
 454 }
 455
 456 show_one(cpuinfo_min_freq, cpuinfo.min_freq);
 457 show_one(cpuinfo_max_freq, cpuinfo.max_freq);
 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", policy->governor->name);
 516         return -EINVAL;
 517 }
 518
 519
 520 /**
 521  * store_scaling_governor - store policy for the specified CPU
 522  */
 523 static ssize_t store_scaling_governor(struct cpufreq_policy *policy,
 524                                         const char *buf, size_t count)
 525 {
 526         unsigned int ret = -EINVAL;
 527         char    str_governor[16];
 528         struct cpufreq_policy new_policy;
 529
 530         ret = cpufreq_get_policy(&new_policy, policy->cpu);
 531         if (ret)
 532                 return ret;
 533
 534         ret = sscanf (buf, "%15s", str_governor);
 535         if (ret != 1)
 536                 return -EINVAL;
 537
 538         if (cpufreq_parse_governor(str_governor, &new_policy.policy,
 539                                                 &new_policy.governor))
 540                 return -EINVAL;
 541
 542         /* Do not use cpufreq_set_policy here or the user_policy.max
 543            will be wrongly overridden */
 544         ret = __cpufreq_set_policy(policy, &new_policy);
 545
 546         policy->user_policy.policy = policy->policy;
 547         policy->user_policy.governor = policy->governor;
 548
 549         if (ret)
 550                 return ret;
 551         else
 552                 return count;
 553 }
 554
 555 /**
 556  * show_scaling_driver - show the cpufreq driver currently loaded
 557  */
 558 static ssize_t show_scaling_driver(struct cpufreq_policy *policy, char *buf)
 559 {
 560         return scnprintf(buf, CPUFREQ_NAME_LEN, "%s\n", cpufreq_driver->name);
 561 }
 562
 563 /**
 564  * show_scaling_available_governors - show the available CPUfreq governors
 565  */
 566 static ssize_t show_scaling_available_governors(struct cpufreq_policy *policy,
 567                                                 char *buf)
 568 {
 569         ssize_t i = 0;
 570         struct cpufreq_governor *t;
 571
 572         if (!cpufreq_driver->target) {
 573                 i += sprintf(buf, "performance powersave");
 574                 goto out;
 575         }
 576
 577         list_for_each_entry(t, &cpufreq_governor_list, governor_list) {
 578                 if (i >= (ssize_t) ((PAGE_SIZE / sizeof(char)) - (CPUFREQ_NAME_LEN + 2)))
 579                         goto out;
 580                 i += scnprintf(&buf[i], CPUFREQ_NAME_LEN, "%s ", t->name);
 581         }
 582 out:
 583         i += sprintf(&buf[i], "\n");
 584         return i;
 585 }
 586
 587 static ssize_t show_cpus(const struct cpumask *mask, char *buf)
 588 {
 589         ssize_t i = 0;
 590         unsigned int cpu;
 591
 592         for_each_cpu(cpu, mask) {
 593                 if (i)
 594                         i += scnprintf(&buf[i], (PAGE_SIZE - i - 2), " ");
 595                 i += scnprintf(&buf[i], (PAGE_SIZE - i - 2), "%u", cpu);
 596                 if (i >= (PAGE_SIZE - 5))
 597                     break;
 598         }
 599         i += sprintf(&buf[i], "\n");
 600         return i;
 601 }
 602
 603 /**
 604  * show_related_cpus - show the CPUs affected by each transition even if
 605  * hw coordination is in use
 606  */
 607 static ssize_t show_related_cpus(struct cpufreq_policy *policy, char *buf)
 608 {
 609         if (cpumask_empty(policy->related_cpus))
 610                 return show_cpus(policy->cpus, buf);
 611         return show_cpus(policy->related_cpus, buf);
 612 }
 613
 614 /**
 615  * show_affected_cpus - show the CPUs affected by each transition
 616  */
 617 static ssize_t show_affected_cpus(struct cpufreq_policy *policy, char *buf)
 618 {
 619         return show_cpus(policy->cpus, buf);
 620 }
 621
 622 static ssize_t store_scaling_setspeed(struct cpufreq_policy *policy,
 623                                         const char *buf, size_t count)
 624 {
 625         unsigned int freq = 0;
 626         unsigned int ret;
 627
 628         if (!policy->governor || !policy->governor->store_setspeed)
 629                 return -EINVAL;
 630
 631         ret = sscanf(buf, "%u", &freq);
 632         if (ret != 1)
 633                 return -EINVAL;
 634
 635         policy->governor->store_setspeed(policy, freq);
 636
 637         return count;
 638 }
 639
 640 static ssize_t show_scaling_setspeed(struct cpufreq_policy *policy, char *buf)
 641 {
 642         if (!policy->governor || !policy->governor->show_setspeed)
 643                 return sprintf(buf, "<unsupported>\n");
 644
 645         return policy->governor->show_setspeed(policy, buf);
 646 }
 647
 648 #define define_one_ro(_name) \
 649 static struct freq_attr _name = \
 650 __ATTR(_name, 0444, show_##_name, NULL)
 651
 652 #define define_one_ro0400(_name) \
 653 static struct freq_attr _name = \
 654 __ATTR(_name, 0400, show_##_name, NULL)
 655
 656 #define define_one_rw(_name) \
 657 static struct freq_attr _name = \
 658 __ATTR(_name, 0644, show_##_name, store_##_name)
 659
 660 define_one_ro0400(cpuinfo_cur_freq);
 661 define_one_ro(cpuinfo_min_freq);
 662 define_one_ro(cpuinfo_max_freq);
 663 define_one_ro(scaling_available_governors);
 664 define_one_ro(scaling_driver);
 665 define_one_ro(scaling_cur_freq);
 666 define_one_ro(related_cpus);
 667 define_one_ro(affected_cpus);
 668 define_one_rw(scaling_min_freq);
 669 define_one_rw(scaling_max_freq);
 670 define_one_rw(scaling_governor);
 671 define_one_rw(scaling_setspeed);
 672
 673 static struct attribute *default_attrs[] = {
 674         &cpuinfo_min_freq.attr,
 675         &cpuinfo_max_freq.attr,
 676         &scaling_min_freq.attr,
 677         &scaling_max_freq.attr,
 678         &affected_cpus.attr,
 679         &related_cpus.attr,
 680         &scaling_governor.attr,
 681         &scaling_driver.attr,
 682         &scaling_available_governors.attr,
 683         &scaling_setspeed.attr,
 684         NULL
 685 };
 686
 687 #define to_policy(k) container_of(k,struct cpufreq_policy,kobj)
 688 #define to_attr(a) container_of(a,struct freq_attr,attr)
 689
 690 static ssize_t show(struct kobject *kobj, struct attribute *attr ,char *buf)
 691 {
 692         struct cpufreq_policy *policy = to_policy(kobj);
 693         struct freq_attr *fattr = to_attr(attr);
 694         ssize_t ret = -EINVAL;
 695         policy = cpufreq_cpu_get(policy->cpu);
 696         if (!policy)
 697                 goto no_policy;
 698
 699         if (lock_policy_rwsem_read(policy->cpu) < 0)
 700                 goto fail;
 701
 702         if (fattr->show)
 703                 ret = fattr->show(policy, buf);
 704         else
 705                 ret = -EIO;
 706
 707         unlock_policy_rwsem_read(policy->cpu);
 708 fail:
 709         cpufreq_cpu_put(policy);
 710 no_policy:
 711         return ret;
 712 }
 713
 714 static ssize_t store(struct kobject *kobj, struct attribute *attr,
 715                      const char *buf, size_t count)
 716 {
 717         struct cpufreq_policy *policy = to_policy(kobj);
 718         struct freq_attr *fattr = to_attr(attr);
 719         ssize_t ret = -EINVAL;
 720         policy = cpufreq_cpu_get(policy->cpu);
 721         if (!policy)
 722                 goto no_policy;
 723
 724         if (lock_policy_rwsem_write(policy->cpu) < 0)
 725                 goto fail;
 726
 727         if (fattr->store)
 728                 ret = fattr->store(policy, buf, count);
 729         else
 730                 ret = -EIO;
 731
 732         unlock_policy_rwsem_write(policy->cpu);
 733 fail:
 734         cpufreq_cpu_put(policy);
 735 no_policy:
 736         return ret;
 737 }
 738
 739 static void cpufreq_sysfs_release(struct kobject *kobj)
 740 {
 741         struct cpufreq_policy *policy = to_policy(kobj);
 742         dprintk("last reference is dropped\n");
 743         complete(&policy->kobj_unregister);
 744 }
 745
 746 static struct sysfs_ops sysfs_ops = {
 747         .show   = show,
 748         .store  = store,
 749 };
 750
 751 static struct kobj_type ktype_cpufreq = {
 752         .sysfs_ops      = &sysfs_ops,
 753         .default_attrs  = default_attrs,
 754         .release        = cpufreq_sysfs_release,
 755 };
 756
 757
 758 /**
 759  * cpufreq_add_dev - add a CPU device
 760  *
 761  * Adds the cpufreq interface for a CPU device.
 762  */
 763 static int cpufreq_add_dev(struct sys_device *sys_dev)
 764 {
 765         unsigned int cpu = sys_dev->id;
 766         int ret = 0;
 767         struct cpufreq_policy new_policy;
 768         struct cpufreq_policy *policy;
 769         struct freq_attr **drv_attr;
 770         struct sys_device *cpu_sys_dev;
 771         unsigned long flags;
 772         unsigned int j;
 773 #ifdef CONFIG_SMP
 774         struct cpufreq_policy *managed_policy;
 775 #endif
 776
 777         if (cpu_is_offline(cpu))
 778                 return 0;
 779
 780         cpufreq_debug_disable_ratelimit();
 781         dprintk("adding CPU %u\n", cpu);
 782
 783 #ifdef CONFIG_SMP
 784         /* check whether a different CPU already registered this
 785          * CPU because it is in the same boat. */
 786         policy = cpufreq_cpu_get(cpu);
 787         if (unlikely(policy)) {
 788                 cpufreq_cpu_put(policy);
 789                 cpufreq_debug_enable_ratelimit();
 790                 return 0;
 791         }
 792 #endif
 793
 794         if (!try_module_get(cpufreq_driver->owner)) {
 795                 ret = -EINVAL;
 796                 goto module_out;
 797         }
 798
 799         policy = kzalloc(sizeof(struct cpufreq_policy), GFP_KERNEL);
 800         if (!policy) {
 801                 ret = -ENOMEM;
 802                 goto nomem_out;
 803         }
 804         if (!alloc_cpumask_var(&policy->cpus, GFP_KERNEL)) {
 805                 kfree(policy);
 806                 ret = -ENOMEM;
 807                 goto nomem_out;
 808         }
 809         if (!alloc_cpumask_var(&policy->related_cpus, GFP_KERNEL)) {
 810                 free_cpumask_var(policy->cpus);
 811                 kfree(policy);
 812                 ret = -ENOMEM;
 813                 goto nomem_out;
 814         }
 815
 816         policy->cpu = cpu;
 817         cpumask_copy(policy->cpus, cpumask_of(cpu));
 818
 819         /* Initially set CPU itself as the policy_cpu */
 820         per_cpu(policy_cpu, cpu) = cpu;
 821         lock_policy_rwsem_write(cpu);
 822
 823         init_completion(&policy->kobj_unregister);
 824         INIT_WORK(&policy->update, handle_update);
 825
 826         /* Set governor before ->init, so that driver could check it */
 827         policy->governor = CPUFREQ_DEFAULT_GOVERNOR;
 828         /* call driver. From then on the cpufreq must be able
 829          * to accept all calls to ->verify and ->setpolicy for this CPU
 830          */
 831         ret = cpufreq_driver->init(policy);
 832         if (ret) {
 833                 dprintk("initialization failed\n");
 834                 goto err_out;
 835         }
 836         policy->user_policy.min = policy->min;
 837         policy->user_policy.max = policy->max;
 838
 839         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
 840                                      CPUFREQ_START, policy);
 841
 842 #ifdef CONFIG_SMP
 843
 844 #ifdef CONFIG_HOTPLUG_CPU
 845         if (per_cpu(cpufreq_cpu_governor, cpu)) {
 846                 policy->governor = per_cpu(cpufreq_cpu_governor, cpu);
 847                 dprintk("Restoring governor %s for cpu %d\n",
 848                        policy->governor->name, cpu);
 849         }
 850 #endif
 851
 852         for_each_cpu(j, policy->cpus) {
 853                 if (cpu == j)
 854                         continue;
 855
 856                 /* check for existing affected CPUs.  They may not be aware
 857                  * of it due to CPU Hotplug.
 858                  */
 859                 managed_policy = cpufreq_cpu_get(j);            // FIXME: Where is this released?  What about error paths?
 860                 if (unlikely(managed_policy)) {
 861
 862                         /* Set proper policy_cpu */
 863                         unlock_policy_rwsem_write(cpu);
 864                         per_cpu(policy_cpu, cpu) = managed_policy->cpu;
 865
 866                         if (lock_policy_rwsem_write(cpu) < 0)
 867                                 goto err_out_driver_exit;
 868
 869                         spin_lock_irqsave(&cpufreq_driver_lock, flags);
 870                         cpumask_copy(managed_policy->cpus, policy->cpus);
 871                         per_cpu(cpufreq_cpu_data, cpu) = managed_policy;
 872                         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
 873
 874                         dprintk("CPU already managed, adding link\n");
 875                         ret = sysfs_create_link(&sys_dev->kobj,
 876                                                 &managed_policy->kobj,
 877                                                 "cpufreq");
 878                         if (ret)
 879                                 goto err_out_driver_exit;
 880
 881                         cpufreq_debug_enable_ratelimit();
 882                         ret = 0;
 883                         goto err_out_driver_exit; /* call driver->exit() */
 884                 }
 885         }
 886 #endif
 887         memcpy(&new_policy, policy, sizeof(struct cpufreq_policy));
 888
 889         /* prepare interface data */
 890         ret = kobject_init_and_add(&policy->kobj, &ktype_cpufreq, &sys_dev->kobj,
 891                                    "cpufreq");
 892         if (ret)
 893                 goto err_out_driver_exit;
 894
 895         /* set up files for this cpu device */
 896         drv_attr = cpufreq_driver->attr;
 897         while ((drv_attr) && (*drv_attr)) {
 898                 ret = sysfs_create_file(&policy->kobj, &((*drv_attr)->attr));
 899                 if (ret)
 900                         goto err_out_driver_exit;
 901                 drv_attr++;
 902         }
 903         if (cpufreq_driver->get) {
 904                 ret = sysfs_create_file(&policy->kobj, &cpuinfo_cur_freq.attr);
 905                 if (ret)
 906                         goto err_out_driver_exit;
 907         }
 908         if (cpufreq_driver->target) {
 909                 ret = sysfs_create_file(&policy->kobj, &scaling_cur_freq.attr);
 910                 if (ret)
 911                         goto err_out_driver_exit;
 912         }
 913
 914         spin_lock_irqsave(&cpufreq_driver_lock, flags);
 915         for_each_cpu(j, policy->cpus) {
 916                 per_cpu(cpufreq_cpu_data, j) = policy;
 917                 per_cpu(policy_cpu, j) = policy->cpu;
 918         }
 919         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
 920
 921         /* symlink affected CPUs */
 922         for_each_cpu(j, policy->cpus) {
 923                 if (j == cpu)
 924                         continue;
 925                 if (!cpu_online(j))
 926                         continue;
 927
 928                 dprintk("CPU %u already managed, adding link\n", j);
 929                 cpufreq_cpu_get(cpu);
 930                 cpu_sys_dev = get_cpu_sysdev(j);
 931                 ret = sysfs_create_link(&cpu_sys_dev->kobj, &policy->kobj,
 932                                         "cpufreq");
 933                 if (ret)
 934                         goto err_out_unregister;
 935         }
 936
 937         policy->governor = NULL; /* to assure that the starting sequence is
 938                                   * run in cpufreq_set_policy */
 939
 940         /* set default policy */
 941         ret = __cpufreq_set_policy(policy, &new_policy);
 942         policy->user_policy.policy = policy->policy;
 943         policy->user_policy.governor = policy->governor;
 944
 945         if (ret) {
 946                 dprintk("setting policy failed\n");
 947                 goto err_out_unregister;
 948         }
 949
 950         unlock_policy_rwsem_write(cpu);
 951
 952         kobject_uevent(&policy->kobj, KOBJ_ADD);
 953         module_put(cpufreq_driver->owner);
 954         dprintk("initialization complete\n");
 955         cpufreq_debug_enable_ratelimit();
 956
 957         return 0;
 958
 959
 960 err_out_unregister:
 961         spin_lock_irqsave(&cpufreq_driver_lock, flags);
 962         for_each_cpu(j, policy->cpus)
 963                 per_cpu(cpufreq_cpu_data, j) = NULL;
 964         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
 965
 966         kobject_put(&policy->kobj);
 967         wait_for_completion(&policy->kobj_unregister);
 968
 969 err_out_driver_exit:
 970         if (cpufreq_driver->exit)
 971                 cpufreq_driver->exit(policy);
 972
 973 err_out:
 974         unlock_policy_rwsem_write(cpu);
 975         kfree(policy);
 976
 977 nomem_out:
 978         module_put(cpufreq_driver->owner);
 979 module_out:
 980         cpufreq_debug_enable_ratelimit();
 981         return ret;
 982 }
 983
 984
 985 /**
 986  * __cpufreq_remove_dev - remove a CPU device
 987  *
 988  * Removes the cpufreq interface for a CPU device.
 989  * Caller should already have policy_rwsem in write mode for this CPU.
 990  * This routine frees the rwsem before returning.
 991  */
 992 static int __cpufreq_remove_dev(struct sys_device *sys_dev)
 993 {
 994         unsigned int cpu = sys_dev->id;
 995         unsigned long flags;
 996         struct cpufreq_policy *data;
 997 #ifdef CONFIG_SMP
 998         struct sys_device *cpu_sys_dev;
 999         unsigned int j;
1000 #endif
1001
1002         cpufreq_debug_disable_ratelimit();
1003         dprintk("unregistering CPU %u\n", cpu);
1004
1005         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1006         data = per_cpu(cpufreq_cpu_data, cpu);
1007
1008         if (!data) {
1009                 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1010                 cpufreq_debug_enable_ratelimit();
1011                 unlock_policy_rwsem_write(cpu);
1012                 return -EINVAL;
1013         }
1014         per_cpu(cpufreq_cpu_data, cpu) = NULL;
1015
1016
1017 #ifdef CONFIG_SMP
1018         /* if this isn't the CPU which is the parent of the kobj, we
1019          * only need to unlink, put and exit
1020          */
1021         if (unlikely(cpu != data->cpu)) {
1022                 dprintk("removing link\n");
1023                 cpumask_clear_cpu(cpu, data->cpus);
1024                 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1025                 sysfs_remove_link(&sys_dev->kobj, "cpufreq");
1026                 cpufreq_cpu_put(data);
1027                 cpufreq_debug_enable_ratelimit();
1028                 unlock_policy_rwsem_write(cpu);
1029                 return 0;
1030         }
1031 #endif
1032
1033 #ifdef CONFIG_SMP
1034
1035 #ifdef CONFIG_HOTPLUG_CPU
1036         per_cpu(cpufreq_cpu_governor, cpu) = data->governor;
1037 #endif
1038
1039         /* if we have other CPUs still registered, we need to unlink them,
1040          * or else wait_for_completion below will lock up. Clean the
1041          * per_cpu(cpufreq_cpu_data) while holding the lock, and remove
1042          * the sysfs links afterwards.
1043          */
1044         if (unlikely(cpumask_weight(data->cpus) > 1)) {
1045                 for_each_cpu(j, data->cpus) {
1046                         if (j == cpu)
1047                                 continue;
1048                         per_cpu(cpufreq_cpu_data, j) = NULL;
1049                 }
1050         }
1051
1052         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1053
1054         if (unlikely(cpumask_weight(data->cpus) > 1)) {
1055                 for_each_cpu(j, data->cpus) {
1056                         if (j == cpu)
1057                                 continue;
1058                         dprintk("removing link for cpu %u\n", j);
1059 #ifdef CONFIG_HOTPLUG_CPU
1060                         per_cpu(cpufreq_cpu_governor, j) = data->governor;
1061 #endif
1062                         cpu_sys_dev = get_cpu_sysdev(j);
1063                         sysfs_remove_link(&cpu_sys_dev->kobj, "cpufreq");
1064                         cpufreq_cpu_put(data);
1065                 }
1066         }
1067 #else
1068         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1069 #endif
1070
1071         if (cpufreq_driver->target)
1072                 __cpufreq_governor(data, CPUFREQ_GOV_STOP);
1073
1074         unlock_policy_rwsem_write(cpu);
1075
1076         kobject_put(&data->kobj);
1077
1078         /* we need to make sure that the underlying kobj is actually
1079          * not referenced anymore by anybody before we proceed with
1080          * unloading.
1081          */
1082         dprintk("waiting for dropping of refcount\n");
1083         wait_for_completion(&data->kobj_unregister);
1084         dprintk("wait complete\n");
1085
1086         if (cpufreq_driver->exit)
1087                 cpufreq_driver->exit(data);
1088
1089         free_cpumask_var(data->related_cpus);
1090         free_cpumask_var(data->cpus);
1091         kfree(data);
1092         per_cpu(cpufreq_cpu_data, cpu) = NULL;
1093
1094         cpufreq_debug_enable_ratelimit();
1095         return 0;
1096 }
1097
1098
1099 static int cpufreq_remove_dev(struct sys_device *sys_dev)
1100 {
1101         unsigned int cpu = sys_dev->id;
1102         int retval;
1103
1104         if (cpu_is_offline(cpu))
1105                 return 0;
1106
1107         if (unlikely(lock_policy_rwsem_write(cpu)))
1108                 BUG();
1109
1110         retval = __cpufreq_remove_dev(sys_dev);
1111         return retval;
1112 }
1113
1114
1115 static void handle_update(struct work_struct *work)
1116 {
1117         struct cpufreq_policy *policy =
1118                 container_of(work, struct cpufreq_policy, update);
1119         unsigned int cpu = policy->cpu;
1120         dprintk("handle_update for cpu %u called\n", cpu);
1121         cpufreq_update_policy(cpu);
1122 }
1123
1124 /**
1125  *      cpufreq_out_of_sync - If actual and saved CPU frequency differs, we're in deep trouble.
1126  *      @cpu: cpu number
1127  *      @old_freq: CPU frequency the kernel thinks the CPU runs at
1128  *      @new_freq: CPU frequency the CPU actually runs at
1129  *
1130  *      We adjust to current frequency first, and need to clean up later. So either call
1131  *      to cpufreq_update_policy() or schedule handle_update()).
1132  */
1133 static void cpufreq_out_of_sync(unsigned int cpu, unsigned int old_freq,
1134                                 unsigned int new_freq)
1135 {
1136         struct cpufreq_freqs freqs;
1137
1138         dprintk("Warning: CPU frequency out of sync: cpufreq and timing "
1139                "core thinks of %u, is %u kHz.\n", old_freq, new_freq);
1140
1141         freqs.cpu = cpu;
1142         freqs.old = old_freq;
1143         freqs.new = new_freq;
1144         cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1145         cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1146 }
1147
1148
1149 /**
1150  * cpufreq_quick_get - get the CPU frequency (in kHz) from policy->cur
1151  * @cpu: CPU number
1152  *
1153  * This is the last known freq, without actually getting it from the driver.
1154  * Return value will be same as what is shown in scaling_cur_freq in sysfs.
1155  */
1156 unsigned int cpufreq_quick_get(unsigned int cpu)
1157 {
1158         struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
1159         unsigned int ret_freq = 0;
1160
1161         if (policy) {
1162                 ret_freq = policy->cur;
1163                 cpufreq_cpu_put(policy);
1164         }
1165
1166         return ret_freq;
1167 }
1168 EXPORT_SYMBOL(cpufreq_quick_get);
1169
1170
1171 static unsigned int __cpufreq_get(unsigned int cpu)
1172 {
1173         struct cpufreq_policy *policy = per_cpu(cpufreq_cpu_data, cpu);
1174         unsigned int ret_freq = 0;
1175
1176         if (!cpufreq_driver->get)
1177                 return ret_freq;
1178
1179         ret_freq = cpufreq_driver->get(cpu);
1180
1181         if (ret_freq && policy->cur &&
1182                 !(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
1183                 /* verify no discrepancy between actual and
1184                                         saved value exists */
1185                 if (unlikely(ret_freq != policy->cur)) {
1186                         cpufreq_out_of_sync(cpu, policy->cur, ret_freq);
1187                         schedule_work(&policy->update);
1188                 }
1189         }
1190
1191         return ret_freq;
1192 }
1193
1194 /**
1195  * cpufreq_get - get the current CPU frequency (in kHz)
1196  * @cpu: CPU number
1197  *
1198  * Get the CPU current (static) CPU frequency
1199  */
1200 unsigned int cpufreq_get(unsigned int cpu)
1201 {
1202         unsigned int ret_freq = 0;
1203         struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
1204
1205         if (!policy)
1206                 goto out;
1207
1208         if (unlikely(lock_policy_rwsem_read(cpu)))
1209                 goto out_policy;
1210
1211         ret_freq = __cpufreq_get(cpu);
1212
1213         unlock_policy_rwsem_read(cpu);
1214
1215 out_policy:
1216         cpufreq_cpu_put(policy);
1217 out:
1218         return ret_freq;
1219 }
1220 EXPORT_SYMBOL(cpufreq_get);
1221
1222
1223 /**
1224  *      cpufreq_suspend - let the low level driver prepare for suspend
1225  */
1226
1227 static int cpufreq_suspend(struct sys_device *sysdev, pm_message_t pmsg)
1228 {
1229         int cpu = sysdev->id;
1230         int ret = 0;
1231         unsigned int cur_freq = 0;
1232         struct cpufreq_policy *cpu_policy;
1233
1234         dprintk("suspending cpu %u\n", cpu);
1235
1236         if (!cpu_online(cpu))
1237                 return 0;
1238
1239         /* we may be lax here as interrupts are off. Nonetheless
1240          * we need to grab the correct cpu policy, as to check
1241          * whether we really run on this CPU.
1242          */
1243
1244         cpu_policy = cpufreq_cpu_get(cpu);
1245         if (!cpu_policy)
1246                 return -EINVAL;
1247
1248         /* only handle each CPU group once */
1249         if (unlikely(cpu_policy->cpu != cpu))
1250                 goto out;
1251
1252         if (cpufreq_driver->suspend) {
1253                 ret = cpufreq_driver->suspend(cpu_policy, pmsg);
1254                 if (ret) {
1255                         printk(KERN_ERR "cpufreq: suspend failed in ->suspend "
1256                                         "step on CPU %u\n", cpu_policy->cpu);
1257                         goto out;
1258                 }
1259         }
1260
1261         if (cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)
1262                 goto out;
1263
1264         if (cpufreq_driver->get)
1265                 cur_freq = cpufreq_driver->get(cpu_policy->cpu);
1266
1267         if (!cur_freq || !cpu_policy->cur) {
1268                 printk(KERN_ERR "cpufreq: suspend failed to assert current "
1269                        "frequency is what timing core thinks it is.\n");
1270                 goto out;
1271         }
1272
1273         if (unlikely(cur_freq != cpu_policy->cur)) {
1274                 struct cpufreq_freqs freqs;
1275
1276                 if (!(cpufreq_driver->flags & CPUFREQ_PM_NO_WARN))
1277                         dprintk("Warning: CPU frequency is %u, "
1278                                "cpufreq assumed %u kHz.\n",
1279                                cur_freq, cpu_policy->cur);
1280
1281                 freqs.cpu = cpu;
1282                 freqs.old = cpu_policy->cur;
1283                 freqs.new = cur_freq;
1284
1285                 srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
1286                                     CPUFREQ_SUSPENDCHANGE, &freqs);
1287                 adjust_jiffies(CPUFREQ_SUSPENDCHANGE, &freqs);
1288
1289                 cpu_policy->cur = cur_freq;
1290         }
1291
1292 out:
1293         cpufreq_cpu_put(cpu_policy);
1294         return ret;
1295 }
1296
1297 /**
1298  *      cpufreq_resume -  restore proper CPU frequency handling after resume
1299  *
1300  *      1.) resume CPUfreq hardware support (cpufreq_driver->resume())
1301  *      2.) if ->target and !CPUFREQ_CONST_LOOPS: verify we're in sync
1302  *      3.) schedule call cpufreq_update_policy() ASAP as interrupts are
1303  *          restored.
1304  */
1305 static int cpufreq_resume(struct sys_device *sysdev)
1306 {
1307         int cpu = sysdev->id;
1308         int ret = 0;
1309         struct cpufreq_policy *cpu_policy;
1310
1311         dprintk("resuming cpu %u\n", cpu);
1312
1313         if (!cpu_online(cpu))
1314                 return 0;
1315
1316         /* we may be lax here as interrupts are off. Nonetheless
1317          * we need to grab the correct cpu policy, as to check
1318          * whether we really run on this CPU.
1319          */
1320
1321         cpu_policy = cpufreq_cpu_get(cpu);
1322         if (!cpu_policy)
1323                 return -EINVAL;
1324
1325         /* only handle each CPU group once */
1326         if (unlikely(cpu_policy->cpu != cpu))
1327                 goto fail;
1328
1329         if (cpufreq_driver->resume) {
1330                 ret = cpufreq_driver->resume(cpu_policy);
1331                 if (ret) {
1332                         printk(KERN_ERR "cpufreq: resume failed in ->resume "
1333                                         "step on CPU %u\n", cpu_policy->cpu);
1334                         goto fail;
1335                 }
1336         }
1337
1338         if (!(cpufreq_driver->flags & CPUFREQ_CONST_LOOPS)) {
1339                 unsigned int cur_freq = 0;
1340
1341                 if (cpufreq_driver->get)
1342                         cur_freq = cpufreq_driver->get(cpu_policy->cpu);
1343
1344                 if (!cur_freq || !cpu_policy->cur) {
1345                         printk(KERN_ERR "cpufreq: resume failed to assert "
1346                                         "current frequency is what timing core "
1347                                         "thinks it is.\n");
1348                         goto out;
1349                 }
1350
1351                 if (unlikely(cur_freq != cpu_policy->cur)) {
1352                         struct cpufreq_freqs freqs;
1353
1354                         if (!(cpufreq_driver->flags & CPUFREQ_PM_NO_WARN))
1355                                 dprintk("Warning: CPU frequency "
1356                                        "is %u, cpufreq assumed %u kHz.\n",
1357                                        cur_freq, cpu_policy->cur);
1358
1359                         freqs.cpu = cpu;
1360                         freqs.old = cpu_policy->cur;
1361                         freqs.new = cur_freq;
1362
1363                         srcu_notifier_call_chain(
1364                                         &cpufreq_transition_notifier_list,
1365                                         CPUFREQ_RESUMECHANGE, &freqs);
1366                         adjust_jiffies(CPUFREQ_RESUMECHANGE, &freqs);
1367
1368                         cpu_policy->cur = cur_freq;
1369                 }
1370         }
1371
1372 out:
1373         schedule_work(&cpu_policy->update);
1374 fail:
1375         cpufreq_cpu_put(cpu_policy);
1376         return ret;
1377 }
1378
1379 static struct sysdev_driver cpufreq_sysdev_driver = {
1380         .add            = cpufreq_add_dev,
1381         .remove         = cpufreq_remove_dev,
1382         .suspend        = cpufreq_suspend,
1383         .resume         = cpufreq_resume,
1384 };
1385
1386
1387 /*********************************************************************
1388  *                     NOTIFIER LISTS INTERFACE                      *
1389  *********************************************************************/
1390
1391 /**
1392  *      cpufreq_register_notifier - register a driver with cpufreq
1393  *      @nb: notifier function to register
1394  *      @list: CPUFREQ_TRANSITION_NOTIFIER or CPUFREQ_POLICY_NOTIFIER
1395  *
1396  *      Add a driver to one of two lists: either a list of drivers that
1397  *      are notified about clock rate changes (once before and once after
1398  *      the transition), or a list of drivers that are notified about
1399  *      changes in cpufreq policy.
1400  *
1401  *      This function may sleep, and has the same return conditions as
1402  *      blocking_notifier_chain_register.
1403  */
1404 int cpufreq_register_notifier(struct notifier_block *nb, unsigned int list)
1405 {
1406         int ret;
1407
1408         WARN_ON(!init_cpufreq_transition_notifier_list_called);
1409
1410         switch (list) {
1411         case CPUFREQ_TRANSITION_NOTIFIER:
1412                 ret = srcu_notifier_chain_register(
1413                                 &cpufreq_transition_notifier_list, nb);
1414                 break;
1415         case CPUFREQ_POLICY_NOTIFIER:
1416                 ret = blocking_notifier_chain_register(
1417                                 &cpufreq_policy_notifier_list, nb);
1418                 break;
1419         default:
1420                 ret = -EINVAL;
1421         }
1422
1423         return ret;
1424 }
1425 EXPORT_SYMBOL(cpufreq_register_notifier);
1426
1427
1428 /**
1429  *      cpufreq_unregister_notifier - unregister a driver with cpufreq
1430  *      @nb: notifier block to be unregistered
1431  *      @list: CPUFREQ_TRANSITION_NOTIFIER or CPUFREQ_POLICY_NOTIFIER
1432  *
1433  *      Remove a driver from the CPU frequency notifier list.
1434  *
1435  *      This function may sleep, and has the same return conditions as
1436  *      blocking_notifier_chain_unregister.
1437  */
1438 int cpufreq_unregister_notifier(struct notifier_block *nb, unsigned int list)
1439 {
1440         int ret;
1441
1442         switch (list) {
1443         case CPUFREQ_TRANSITION_NOTIFIER:
1444                 ret = srcu_notifier_chain_unregister(
1445                                 &cpufreq_transition_notifier_list, nb);
1446                 break;
1447         case CPUFREQ_POLICY_NOTIFIER:
1448                 ret = blocking_notifier_chain_unregister(
1449                                 &cpufreq_policy_notifier_list, nb);
1450                 break;
1451         default:
1452                 ret = -EINVAL;
1453         }
1454
1455         return ret;
1456 }
1457 EXPORT_SYMBOL(cpufreq_unregister_notifier);
1458
1459
1460 /*********************************************************************
1461  *                              GOVERNORS                            *
1462  *********************************************************************/
1463
1464
1465 int __cpufreq_driver_target(struct cpufreq_policy *policy,
1466                             unsigned int target_freq,
1467                             unsigned int relation)
1468 {
1469         int retval = -EINVAL;
1470
1471         dprintk("target for CPU %u: %u kHz, relation %u\n", policy->cpu,
1472                 target_freq, relation);
1473         if (cpu_online(policy->cpu) && cpufreq_driver->target)
1474                 retval = cpufreq_driver->target(policy, target_freq, relation);
1475
1476         return retval;
1477 }
1478 EXPORT_SYMBOL_GPL(__cpufreq_driver_target);
1479
1480 int cpufreq_driver_target(struct cpufreq_policy *policy,
1481                           unsigned int target_freq,
1482                           unsigned int relation)
1483 {
1484         int ret = -EINVAL;
1485
1486         policy = cpufreq_cpu_get(policy->cpu);
1487         if (!policy)
1488                 goto no_policy;
1489
1490         if (unlikely(lock_policy_rwsem_write(policy->cpu)))
1491                 goto fail;
1492
1493         ret = __cpufreq_driver_target(policy, target_freq, relation);
1494
1495         unlock_policy_rwsem_write(policy->cpu);
1496
1497 fail:
1498         cpufreq_cpu_put(policy);
1499 no_policy:
1500         return ret;
1501 }
1502 EXPORT_SYMBOL_GPL(cpufreq_driver_target);
1503
1504 int __cpufreq_driver_getavg(struct cpufreq_policy *policy, unsigned int cpu)
1505 {
1506         int ret = 0;
1507
1508         policy = cpufreq_cpu_get(policy->cpu);
1509         if (!policy)
1510                 return -EINVAL;
1511
1512         if (cpu_online(cpu) && cpufreq_driver->getavg)
1513                 ret = cpufreq_driver->getavg(policy, cpu);
1514
1515         cpufreq_cpu_put(policy);
1516         return ret;
1517 }
1518 EXPORT_SYMBOL_GPL(__cpufreq_driver_getavg);
1519
1520 /*
1521  * when "event" is CPUFREQ_GOV_LIMITS
1522  */
1523
1524 static int __cpufreq_governor(struct cpufreq_policy *policy,
1525                                         unsigned int event)
1526 {
1527         int ret;
1528
1529         /* Only must be defined when default governor is known to have latency
1530            restrictions, like e.g. conservative or ondemand.
1531            That this is the case is already ensured in Kconfig
1532         */
1533 #ifdef CONFIG_CPU_FREQ_GOV_PERFORMANCE
1534         struct cpufreq_governor *gov = &cpufreq_gov_performance;
1535 #else
1536         struct cpufreq_governor *gov = NULL;
1537 #endif
1538
1539         if (policy->governor->max_transition_latency &&
1540             policy->cpuinfo.transition_latency >
1541             policy->governor->max_transition_latency) {
1542                 if (!gov)
1543                         return -EINVAL;
1544                 else {
1545                         printk(KERN_WARNING "%s governor failed, too long"
1546                                " transition latency of HW, fallback"
1547                                " to %s governor\n",
1548                                policy->governor->name,
1549                                gov->name);
1550                         policy->governor = gov;
1551                 }
1552         }
1553
1554         if (!try_module_get(policy->governor->owner))
1555                 return -EINVAL;
1556
1557         dprintk("__cpufreq_governor for CPU %u, event %u\n",
1558                                                 policy->cpu, event);
1559         ret = policy->governor->governor(policy, event);
1560
1561         /* we keep one module reference alive for
1562                         each CPU governed by this CPU */
1563         if ((event != CPUFREQ_GOV_START) || ret)
1564                 module_put(policy->governor->owner);
1565         if ((event == CPUFREQ_GOV_STOP) && !ret)
1566                 module_put(policy->governor->owner);
1567
1568         return ret;
1569 }
1570
1571
1572 int cpufreq_register_governor(struct cpufreq_governor *governor)
1573 {
1574         int err;
1575
1576         if (!governor)
1577                 return -EINVAL;
1578
1579         mutex_lock(&cpufreq_governor_mutex);
1580
1581         err = -EBUSY;
1582         if (__find_governor(governor->name) == NULL) {
1583                 err = 0;
1584                 list_add(&governor->governor_list, &cpufreq_governor_list);
1585         }
1586
1587         mutex_unlock(&cpufreq_governor_mutex);
1588         return err;
1589 }
1590 EXPORT_SYMBOL_GPL(cpufreq_register_governor);
1591
1592
1593 void cpufreq_unregister_governor(struct cpufreq_governor *governor)
1594 {
1595         if (!governor)
1596                 return;
1597
1598         mutex_lock(&cpufreq_governor_mutex);
1599         list_del(&governor->governor_list);
1600         mutex_unlock(&cpufreq_governor_mutex);
1601         return;
1602 }
1603 EXPORT_SYMBOL_GPL(cpufreq_unregister_governor);
1604
1605
1606
1607 /*********************************************************************
1608  *                          POLICY INTERFACE                         *
1609  *********************************************************************/
1610
1611 /**
1612  * cpufreq_get_policy - get the current cpufreq_policy
1613  * @policy: struct cpufreq_policy into which the current cpufreq_policy is written
1614  *
1615  * Reads the current cpufreq policy.
1616  */
1617 int cpufreq_get_policy(struct cpufreq_policy *policy, unsigned int cpu)
1618 {
1619         struct cpufreq_policy *cpu_policy;
1620         if (!policy)
1621                 return -EINVAL;
1622
1623         cpu_policy = cpufreq_cpu_get(cpu);
1624         if (!cpu_policy)
1625                 return -EINVAL;
1626
1627         memcpy(policy, cpu_policy, sizeof(struct cpufreq_policy));
1628
1629         cpufreq_cpu_put(cpu_policy);
1630         return 0;
1631 }
1632 EXPORT_SYMBOL(cpufreq_get_policy);
1633
1634
1635 /*
1636  * data   : current policy.
1637  * policy : policy to be set.
1638  */
1639 static int __cpufreq_set_policy(struct cpufreq_policy *data,
1640                                 struct cpufreq_policy *policy)
1641 {
1642         int ret = 0;
1643
1644         cpufreq_debug_disable_ratelimit();
1645         dprintk("setting new policy for CPU %u: %u - %u kHz\n", policy->cpu,
1646                 policy->min, policy->max);
1647
1648         memcpy(&policy->cpuinfo, &data->cpuinfo,
1649                                 sizeof(struct cpufreq_cpuinfo));
1650
1651         if (policy->min > data->max || policy->max < data->min) {
1652                 ret = -EINVAL;
1653                 goto error_out;
1654         }
1655
1656         /* verify the cpu speed can be set within this limit */
1657         ret = cpufreq_driver->verify(policy);
1658         if (ret)
1659                 goto error_out;
1660
1661         /* adjust if necessary - all reasons */
1662         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1663                         CPUFREQ_ADJUST, policy);
1664
1665         /* adjust if necessary - hardware incompatibility*/
1666         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1667                         CPUFREQ_INCOMPATIBLE, policy);
1668
1669         /* verify the cpu speed can be set within this limit,
1670            which might be different to the first one */
1671         ret = cpufreq_driver->verify(policy);
1672         if (ret)
1673                 goto error_out;
1674
1675         /* notification of the new policy */
1676         blocking_notifier_call_chain(&cpufreq_policy_notifier_list,
1677                         CPUFREQ_NOTIFY, policy);
1678
1679         data->min = policy->min;
1680         data->max = policy->max;
1681
1682         dprintk("new min and max freqs are %u - %u kHz\n",
1683                                         data->min, data->max);
1684
1685         if (cpufreq_driver->setpolicy) {
1686                 data->policy = policy->policy;
1687                 dprintk("setting range\n");
1688                 ret = cpufreq_driver->setpolicy(policy);
1689         } else {
1690                 if (policy->governor != data->governor) {
1691                         /* save old, working values */
1692                         struct cpufreq_governor *old_gov = data->governor;
1693
1694                         dprintk("governor switch\n");
1695
1696                         /* end old governor */
1697                         if (data->governor)
1698                                 __cpufreq_governor(data, CPUFREQ_GOV_STOP);
1699
1700                         /* start new governor */
1701                         data->governor = policy->governor;
1702                         if (__cpufreq_governor(data, CPUFREQ_GOV_START)) {
1703                                 /* new governor failed, so re-start old one */
1704                                 dprintk("starting governor %s failed\n",
1705                                                         data->governor->name);
1706                                 if (old_gov) {
1707                                         data->governor = old_gov;
1708                                         __cpufreq_governor(data,
1709                                                            CPUFREQ_GOV_START);
1710                                 }
1711                                 ret = -EINVAL;
1712                                 goto error_out;
1713                         }
1714                         /* might be a policy change, too, so fall through */
1715                 }
1716                 dprintk("governor: change or update limits\n");
1717                 __cpufreq_governor(data, CPUFREQ_GOV_LIMITS);
1718         }
1719
1720 error_out:
1721         cpufreq_debug_enable_ratelimit();
1722         return ret;
1723 }
1724
1725 /**
1726  *      cpufreq_update_policy - re-evaluate an existing cpufreq policy
1727  *      @cpu: CPU which shall be re-evaluated
1728  *
1729  *      Usefull for policy notifiers which have different necessities
1730  *      at different times.
1731  */
1732 int cpufreq_update_policy(unsigned int cpu)
1733 {
1734         struct cpufreq_policy *data = cpufreq_cpu_get(cpu);
1735         struct cpufreq_policy policy;
1736         int ret;
1737
1738         if (!data) {
1739                 ret = -ENODEV;
1740                 goto no_policy;
1741         }
1742
1743         if (unlikely(lock_policy_rwsem_write(cpu))) {
1744                 ret = -EINVAL;
1745                 goto fail;
1746         }
1747
1748         dprintk("updating policy for CPU %u\n", cpu);
1749         memcpy(&policy, data, sizeof(struct cpufreq_policy));
1750         policy.min = data->user_policy.min;
1751         policy.max = data->user_policy.max;
1752         policy.policy = data->user_policy.policy;
1753         policy.governor = data->user_policy.governor;
1754
1755         /* BIOS might change freq behind our back
1756           -> ask driver for current freq and notify governors about a change */
1757         if (cpufreq_driver->get) {
1758                 policy.cur = cpufreq_driver->get(cpu);
1759                 if (!data->cur) {
1760                         dprintk("Driver did not initialize current freq");
1761                         data->cur = policy.cur;
1762                 } else {
1763                         if (data->cur != policy.cur)
1764                                 cpufreq_out_of_sync(cpu, data->cur,
1765                                                                 policy.cur);
1766                 }
1767         }
1768
1769         ret = __cpufreq_set_policy(data, &policy);
1770
1771         unlock_policy_rwsem_write(cpu);
1772
1773 fail:
1774         cpufreq_cpu_put(data);
1775 no_policy:
1776         return ret;
1777 }
1778 EXPORT_SYMBOL(cpufreq_update_policy);
1779
1780 static int __cpuinit cpufreq_cpu_callback(struct notifier_block *nfb,
1781                                         unsigned long action, void *hcpu)
1782 {
1783         unsigned int cpu = (unsigned long)hcpu;
1784         struct sys_device *sys_dev;
1785
1786         sys_dev = get_cpu_sysdev(cpu);
1787         if (sys_dev) {
1788                 switch (action) {
1789                 case CPU_ONLINE:
1790                 case CPU_ONLINE_FROZEN:
1791                         cpufreq_add_dev(sys_dev);
1792                         break;
1793                 case CPU_DOWN_PREPARE:
1794                 case CPU_DOWN_PREPARE_FROZEN:
1795                         if (unlikely(lock_policy_rwsem_write(cpu)))
1796                                 BUG();
1797
1798                         __cpufreq_remove_dev(sys_dev);
1799                         break;
1800                 case CPU_DOWN_FAILED:
1801                 case CPU_DOWN_FAILED_FROZEN:
1802                         cpufreq_add_dev(sys_dev);
1803                         break;
1804                 }
1805         }
1806         return NOTIFY_OK;
1807 }
1808
1809 static struct notifier_block __refdata cpufreq_cpu_notifier =
1810 {
1811     .notifier_call = cpufreq_cpu_callback,
1812 };
1813
1814 /*********************************************************************
1815  *               REGISTER / UNREGISTER CPUFREQ DRIVER                *
1816  *********************************************************************/
1817
1818 /**
1819  * cpufreq_register_driver - register a CPU Frequency driver
1820  * @driver_data: A struct cpufreq_driver containing the values#
1821  * submitted by the CPU Frequency driver.
1822  *
1823  *   Registers a CPU Frequency driver to this core code. This code
1824  * returns zero on success, -EBUSY when another driver got here first
1825  * (and isn't unregistered in the meantime).
1826  *
1827  */
1828 int cpufreq_register_driver(struct cpufreq_driver *driver_data)
1829 {
1830         unsigned long flags;
1831         int ret;
1832
1833         if (!driver_data || !driver_data->verify || !driver_data->init ||
1834             ((!driver_data->setpolicy) && (!driver_data->target)))
1835                 return -EINVAL;
1836
1837         dprintk("trying to register driver %s\n", driver_data->name);
1838
1839         if (driver_data->setpolicy)
1840                 driver_data->flags |= CPUFREQ_CONST_LOOPS;
1841
1842         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1843         if (cpufreq_driver) {
1844                 spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1845                 return -EBUSY;
1846         }
1847         cpufreq_driver = driver_data;
1848         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1849
1850         ret = sysdev_driver_register(&cpu_sysdev_class,
1851                                         &cpufreq_sysdev_driver);
1852
1853         if ((!ret) && !(cpufreq_driver->flags & CPUFREQ_STICKY)) {
1854                 int i;
1855                 ret = -ENODEV;
1856
1857                 /* check for at least one working CPU */
1858                 for (i = 0; i < nr_cpu_ids; i++)
1859                         if (cpu_possible(i) && per_cpu(cpufreq_cpu_data, i)) {
1860                                 ret = 0;
1861                                 break;
1862                         }
1863
1864                 /* if all ->init() calls failed, unregister */
1865                 if (ret) {
1866                         dprintk("no CPU initialized for driver %s\n",
1867                                                         driver_data->name);
1868                         sysdev_driver_unregister(&cpu_sysdev_class,
1869                                                 &cpufreq_sysdev_driver);
1870
1871                         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1872                         cpufreq_driver = NULL;
1873                         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1874                 }
1875         }
1876
1877         if (!ret) {
1878                 register_hotcpu_notifier(&cpufreq_cpu_notifier);
1879                 dprintk("driver %s up and running\n", driver_data->name);
1880                 cpufreq_debug_enable_ratelimit();
1881         }
1882
1883         return ret;
1884 }
1885 EXPORT_SYMBOL_GPL(cpufreq_register_driver);
1886
1887
1888 /**
1889  * cpufreq_unregister_driver - unregister the current CPUFreq driver
1890  *
1891  *    Unregister the current CPUFreq driver. Only call this if you have
1892  * the right to do so, i.e. if you have succeeded in initialising before!
1893  * Returns zero if successful, and -EINVAL if the cpufreq_driver is
1894  * currently not initialised.
1895  */
1896 int cpufreq_unregister_driver(struct cpufreq_driver *driver)
1897 {
1898         unsigned long flags;
1899
1900         cpufreq_debug_disable_ratelimit();
1901
1902         if (!cpufreq_driver || (driver != cpufreq_driver)) {
1903                 cpufreq_debug_enable_ratelimit();
1904                 return -EINVAL;
1905         }
1906
1907         dprintk("unregistering driver %s\n", driver->name);
1908
1909         sysdev_driver_unregister(&cpu_sysdev_class, &cpufreq_sysdev_driver);
1910         unregister_hotcpu_notifier(&cpufreq_cpu_notifier);
1911
1912         spin_lock_irqsave(&cpufreq_driver_lock, flags);
1913         cpufreq_driver = NULL;
1914         spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
1915
1916         return 0;
1917 }
1918 EXPORT_SYMBOL_GPL(cpufreq_unregister_driver);
1919
1920 static int __init cpufreq_core_init(void)
1921 {
1922         int cpu;
1923
1924         for_each_possible_cpu(cpu) {
1925                 per_cpu(policy_cpu, cpu) = -1;
1926                 init_rwsem(&per_cpu(cpu_policy_rwsem, cpu));
1927         }
1928         return 0;
1929 }
1930
1931 core_initcall(cpufreq_core_init);