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