drivers/cpuidle/cpuidle.c

   1 /*
   2  * cpuidle.c - core cpuidle infrastructure
   3  *
   4  * (C) 2006-2007 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
   5  *               Shaohua Li <shaohua.li@intel.com>
   6  *               Adam Belay <abelay@novell.com>
   7  *
   8  * This code is licenced under the GPL.
   9  */
  10
  11 #include <linux/clockchips.h>
  12 #include <linux/kernel.h>
  13 #include <linux/mutex.h>
  14 #include <linux/sched.h>
  15 #include <linux/notifier.h>
  16 #include <linux/pm_qos.h>
  17 #include <linux/cpu.h>
  18 #include <linux/cpuidle.h>
  19 #include <linux/ktime.h>
  20 #include <linux/hrtimer.h>
  21 #include <linux/module.h>
  22 #include <linux/suspend.h>
  23 #include <linux/tick.h>
  24 #include <trace/events/power.h>
  25
  26 #include "cpuidle.h"
  27
  28 DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
  29 DEFINE_PER_CPU(struct cpuidle_device, cpuidle_dev);
  30
  31 DEFINE_MUTEX(cpuidle_lock);
  32 LIST_HEAD(cpuidle_detected_devices);
  33
  34 static int enabled_devices;
  35 static int off __read_mostly;
  36 static int initialized __read_mostly;
  37
  38 int cpuidle_disabled(void)
  39 {
  40         return off;
  41 }
  42 void disable_cpuidle(void)
  43 {
  44         off = 1;
  45 }
  46
  47 /**
  48  * cpuidle_play_dead - cpu off-lining
  49  *
  50  * Returns in case of an error or no driver
  51  */
  52 int cpuidle_play_dead(void)
  53 {
  54         struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
  55         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
  56         int i;
  57
  58         if (!drv)
  59                 return -ENODEV;
  60
  61         /* Find lowest-power state that supports long-term idle */
  62         for (i = drv->state_count - 1; i >= CPUIDLE_DRIVER_STATE_START; i--)
  63                 if (drv->states[i].enter_dead)
  64                         return drv->states[i].enter_dead(dev, i);
  65
  66         return -ENODEV;
  67 }
  68
  69 /**
  70  * cpuidle_find_deepest_state - Find deepest state meeting specific conditions.
  71  * @drv: cpuidle driver for the given CPU.
  72  * @dev: cpuidle device for the given CPU.
  73  * @freeze: Whether or not the state should be suitable for suspend-to-idle.
  74  */
  75 static int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
  76                                       struct cpuidle_device *dev, bool freeze)
  77 {
  78         unsigned int latency_req = 0;
  79         int i, ret = freeze ? -1 : CPUIDLE_DRIVER_STATE_START - 1;
  80
  81         for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
  82                 struct cpuidle_state *s = &drv->states[i];
  83                 struct cpuidle_state_usage *su = &dev->states_usage[i];
  84
  85                 if (s->disabled || su->disable || s->exit_latency <= latency_req
  86                     || (freeze && !s->enter_freeze))
  87                         continue;
  88
  89                 latency_req = s->exit_latency;
  90                 ret = i;
  91         }
  92         return ret;
  93 }
  94
  95 static void enter_freeze_proper(struct cpuidle_driver *drv,
  96                                 struct cpuidle_device *dev, int index)
  97 {
  98         tick_freeze();
  99         /*
 100          * The state used here cannot be a "coupled" one, because the "coupled"
 101          * cpuidle mechanism enables interrupts and doing that with timekeeping
 102          * suspended is generally unsafe.
 103          */
 104         drv->states[index].enter_freeze(dev, drv, index);
 105         WARN_ON(!irqs_disabled());
 106         /*
 107          * timekeeping_resume() that will be called by tick_unfreeze() for the
 108          * last CPU executing it calls functions containing RCU read-side
 109          * critical sections, so tell RCU about that.
 110          */
 111         RCU_NONIDLE(tick_unfreeze());
 112 }
 113
 114 /**
 115  * cpuidle_enter_freeze - Enter an idle state suitable for suspend-to-idle.
 116  *
 117  * If there are states with the ->enter_freeze callback, find the deepest of
 118  * them and enter it with frozen tick.  Otherwise, find the deepest state
 119  * available and enter it normally.
 120  */
 121 void cpuidle_enter_freeze(void)
 122 {
 123         struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
 124         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 125         int index;
 126
 127         /*
 128          * Find the deepest state with ->enter_freeze present, which guarantees
 129          * that interrupts won't be enabled when it exits and allows the tick to
 130          * be frozen safely.
 131          */
 132         index = cpuidle_find_deepest_state(drv, dev, true);
 133         if (index >= 0) {
 134                 enter_freeze_proper(drv, dev, index);
 135                 return;
 136         }
 137
 138         /*
 139          * It is not safe to freeze the tick, find the deepest state available
 140          * at all and try to enter it normally.
 141          */
 142         index = cpuidle_find_deepest_state(drv, dev, false);
 143         if (index >= 0)
 144                 cpuidle_enter(drv, dev, index);
 145         else
 146                 arch_cpu_idle();
 147
 148         /* Interrupts are enabled again here. */
 149         local_irq_disable();
 150 }
 151
 152 /**
 153  * cpuidle_enter_state - enter the state and update stats
 154  * @dev: cpuidle device for this cpu
 155  * @drv: cpuidle driver for this cpu
 156  * @next_state: index into drv->states of the state to enter
 157  */
 158 int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
 159                         int index)
 160 {
 161         int entered_state;
 162
 163         struct cpuidle_state *target_state = &drv->states[index];
 164         ktime_t time_start, time_end;
 165         s64 diff;
 166
 167         trace_cpu_idle_rcuidle(index, dev->cpu);
 168         time_start = ktime_get();
 169
 170         entered_state = target_state->enter(dev, drv, index);
 171
 172         time_end = ktime_get();
 173         trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
 174
 175         if (!cpuidle_state_is_coupled(dev, drv, entered_state))
 176                 local_irq_enable();
 177
 178         diff = ktime_to_us(ktime_sub(time_end, time_start));
 179         if (diff > INT_MAX)
 180                 diff = INT_MAX;
 181
 182         dev->last_residency = (int) diff;
 183
 184         if (entered_state >= 0) {
 185                 /* Update cpuidle counters */
 186                 /* This can be moved to within driver enter routine
 187                  * but that results in multiple copies of same code.
 188                  */
 189                 dev->states_usage[entered_state].time += dev->last_residency;
 190                 dev->states_usage[entered_state].usage++;
 191         } else {
 192                 dev->last_residency = 0;
 193         }
 194
 195         return entered_state;
 196 }
 197
 198 /**
 199  * cpuidle_select - ask the cpuidle framework to choose an idle state
 200  *
 201  * @drv: the cpuidle driver
 202  * @dev: the cpuidle device
 203  *
 204  * Returns the index of the idle state.
 205  */
 206 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
 207 {
 208         if (off || !initialized)
 209                 return -ENODEV;
 210
 211         if (!drv || !dev || !dev->enabled)
 212                 return -EBUSY;
 213
 214         return cpuidle_curr_governor->select(drv, dev);
 215 }
 216
 217 /**
 218  * cpuidle_enter - enter into the specified idle state
 219  *
 220  * @drv:   the cpuidle driver tied with the cpu
 221  * @dev:   the cpuidle device
 222  * @index: the index in the idle state table
 223  *
 224  * Returns the index in the idle state, < 0 in case of error.
 225  * The error code depends on the backend driver
 226  */
 227 int cpuidle_enter(struct cpuidle_driver *drv, struct cpuidle_device *dev,
 228                   int index)
 229 {
 230         if (cpuidle_state_is_coupled(dev, drv, index))
 231                 return cpuidle_enter_state_coupled(dev, drv, index);
 232         return cpuidle_enter_state(dev, drv, index);
 233 }
 234
 235 /**
 236  * cpuidle_reflect - tell the underlying governor what was the state
 237  * we were in
 238  *
 239  * @dev  : the cpuidle device
 240  * @index: the index in the idle state table
 241  *
 242  */
 243 void cpuidle_reflect(struct cpuidle_device *dev, int index)
 244 {
 245         if (cpuidle_curr_governor->reflect)
 246                 cpuidle_curr_governor->reflect(dev, index);
 247 }
 248
 249 /**
 250  * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
 251  */
 252 void cpuidle_install_idle_handler(void)
 253 {
 254         if (enabled_devices) {
 255                 /* Make sure all changes finished before we switch to new idle */
 256                 smp_wmb();
 257                 initialized = 1;
 258         }
 259 }
 260
 261 /**
 262  * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
 263  */
 264 void cpuidle_uninstall_idle_handler(void)
 265 {
 266         if (enabled_devices) {
 267                 initialized = 0;
 268                 wake_up_all_idle_cpus();
 269         }
 270
 271         /*
 272          * Make sure external observers (such as the scheduler)
 273          * are done looking at pointed idle states.
 274          */
 275         synchronize_rcu();
 276 }
 277
 278 /**
 279  * cpuidle_pause_and_lock - temporarily disables CPUIDLE
 280  */
 281 void cpuidle_pause_and_lock(void)
 282 {
 283         mutex_lock(&cpuidle_lock);
 284         cpuidle_uninstall_idle_handler();
 285 }
 286
 287 EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
 288
 289 /**
 290  * cpuidle_resume_and_unlock - resumes CPUIDLE operation
 291  */
 292 void cpuidle_resume_and_unlock(void)
 293 {
 294         cpuidle_install_idle_handler();
 295         mutex_unlock(&cpuidle_lock);
 296 }
 297
 298 EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
 299
 300 /* Currently used in suspend/resume path to suspend cpuidle */
 301 void cpuidle_pause(void)
 302 {
 303         mutex_lock(&cpuidle_lock);
 304         cpuidle_uninstall_idle_handler();
 305         mutex_unlock(&cpuidle_lock);
 306 }
 307
 308 /* Currently used in suspend/resume path to resume cpuidle */
 309 void cpuidle_resume(void)
 310 {
 311         mutex_lock(&cpuidle_lock);
 312         cpuidle_install_idle_handler();
 313         mutex_unlock(&cpuidle_lock);
 314 }
 315
 316 /**
 317  * cpuidle_enable_device - enables idle PM for a CPU
 318  * @dev: the CPU
 319  *
 320  * This function must be called between cpuidle_pause_and_lock and
 321  * cpuidle_resume_and_unlock when used externally.
 322  */
 323 int cpuidle_enable_device(struct cpuidle_device *dev)
 324 {
 325         int ret;
 326         struct cpuidle_driver *drv;
 327
 328         if (!dev)
 329                 return -EINVAL;
 330
 331         if (dev->enabled)
 332                 return 0;
 333
 334         drv = cpuidle_get_cpu_driver(dev);
 335
 336         if (!drv || !cpuidle_curr_governor)
 337                 return -EIO;
 338
 339         if (!dev->registered)
 340                 return -EINVAL;
 341
 342         if (!dev->state_count)
 343                 dev->state_count = drv->state_count;
 344
 345         ret = cpuidle_add_device_sysfs(dev);
 346         if (ret)
 347                 return ret;
 348
 349         if (cpuidle_curr_governor->enable &&
 350             (ret = cpuidle_curr_governor->enable(drv, dev)))
 351                 goto fail_sysfs;
 352
 353         smp_wmb();
 354
 355         dev->enabled = 1;
 356
 357         enabled_devices++;
 358         return 0;
 359
 360 fail_sysfs:
 361         cpuidle_remove_device_sysfs(dev);
 362
 363         return ret;
 364 }
 365
 366 EXPORT_SYMBOL_GPL(cpuidle_enable_device);
 367
 368 /**
 369  * cpuidle_disable_device - disables idle PM for a CPU
 370  * @dev: the CPU
 371  *
 372  * This function must be called between cpuidle_pause_and_lock and
 373  * cpuidle_resume_and_unlock when used externally.
 374  */
 375 void cpuidle_disable_device(struct cpuidle_device *dev)
 376 {
 377         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 378
 379         if (!dev || !dev->enabled)
 380                 return;
 381
 382         if (!drv || !cpuidle_curr_governor)
 383                 return;
 384
 385         dev->enabled = 0;
 386
 387         if (cpuidle_curr_governor->disable)
 388                 cpuidle_curr_governor->disable(drv, dev);
 389
 390         cpuidle_remove_device_sysfs(dev);
 391         enabled_devices--;
 392 }
 393
 394 EXPORT_SYMBOL_GPL(cpuidle_disable_device);
 395
 396 static void __cpuidle_unregister_device(struct cpuidle_device *dev)
 397 {
 398         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 399
 400         list_del(&dev->device_list);
 401         per_cpu(cpuidle_devices, dev->cpu) = NULL;
 402         module_put(drv->owner);
 403 }
 404
 405 static void __cpuidle_device_init(struct cpuidle_device *dev)
 406 {
 407         memset(dev->states_usage, 0, sizeof(dev->states_usage));
 408         dev->last_residency = 0;
 409 }
 410
 411 /**
 412  * __cpuidle_register_device - internal register function called before register
 413  * and enable routines
 414  * @dev: the cpu
 415  *
 416  * cpuidle_lock mutex must be held before this is called
 417  */
 418 static int __cpuidle_register_device(struct cpuidle_device *dev)
 419 {
 420         int ret;
 421         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 422
 423         if (!try_module_get(drv->owner))
 424                 return -EINVAL;
 425
 426         per_cpu(cpuidle_devices, dev->cpu) = dev;
 427         list_add(&dev->device_list, &cpuidle_detected_devices);
 428
 429         ret = cpuidle_coupled_register_device(dev);
 430         if (ret)
 431                 __cpuidle_unregister_device(dev);
 432         else
 433                 dev->registered = 1;
 434
 435         return ret;
 436 }
 437
 438 /**
 439  * cpuidle_register_device - registers a CPU's idle PM feature
 440  * @dev: the cpu
 441  */
 442 int cpuidle_register_device(struct cpuidle_device *dev)
 443 {
 444         int ret = -EBUSY;
 445
 446         if (!dev)
 447                 return -EINVAL;
 448
 449         mutex_lock(&cpuidle_lock);
 450
 451         if (dev->registered)
 452                 goto out_unlock;
 453
 454         __cpuidle_device_init(dev);
 455
 456         ret = __cpuidle_register_device(dev);
 457         if (ret)
 458                 goto out_unlock;
 459
 460         ret = cpuidle_add_sysfs(dev);
 461         if (ret)
 462                 goto out_unregister;
 463
 464         ret = cpuidle_enable_device(dev);
 465         if (ret)
 466                 goto out_sysfs;
 467
 468         cpuidle_install_idle_handler();
 469
 470 out_unlock:
 471         mutex_unlock(&cpuidle_lock);
 472
 473         return ret;
 474
 475 out_sysfs:
 476         cpuidle_remove_sysfs(dev);
 477 out_unregister:
 478         __cpuidle_unregister_device(dev);
 479         goto out_unlock;
 480 }
 481
 482 EXPORT_SYMBOL_GPL(cpuidle_register_device);
 483
 484 /**
 485  * cpuidle_unregister_device - unregisters a CPU's idle PM feature
 486  * @dev: the cpu
 487  */
 488 void cpuidle_unregister_device(struct cpuidle_device *dev)
 489 {
 490         if (!dev || dev->registered == 0)
 491                 return;
 492
 493         cpuidle_pause_and_lock();
 494
 495         cpuidle_disable_device(dev);
 496
 497         cpuidle_remove_sysfs(dev);
 498
 499         __cpuidle_unregister_device(dev);
 500
 501         cpuidle_coupled_unregister_device(dev);
 502
 503         cpuidle_resume_and_unlock();
 504 }
 505
 506 EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
 507
 508 /**
 509  * cpuidle_unregister: unregister a driver and the devices. This function
 510  * can be used only if the driver has been previously registered through
 511  * the cpuidle_register function.
 512  *
 513  * @drv: a valid pointer to a struct cpuidle_driver
 514  */
 515 void cpuidle_unregister(struct cpuidle_driver *drv)
 516 {
 517         int cpu;
 518         struct cpuidle_device *device;
 519
 520         for_each_cpu(cpu, drv->cpumask) {
 521                 device = &per_cpu(cpuidle_dev, cpu);
 522                 cpuidle_unregister_device(device);
 523         }
 524
 525         cpuidle_unregister_driver(drv);
 526 }
 527 EXPORT_SYMBOL_GPL(cpuidle_unregister);
 528
 529 /**
 530  * cpuidle_register: registers the driver and the cpu devices with the
 531  * coupled_cpus passed as parameter. This function is used for all common
 532  * initialization pattern there are in the arch specific drivers. The
 533  * devices is globally defined in this file.
 534  *
 535  * @drv         : a valid pointer to a struct cpuidle_driver
 536  * @coupled_cpus: a cpumask for the coupled states
 537  *
 538  * Returns 0 on success, < 0 otherwise
 539  */
 540 int cpuidle_register(struct cpuidle_driver *drv,
 541                      const struct cpumask *const coupled_cpus)
 542 {
 543         int ret, cpu;
 544         struct cpuidle_device *device;
 545
 546         ret = cpuidle_register_driver(drv);
 547         if (ret) {
 548                 pr_err("failed to register cpuidle driver\n");
 549                 return ret;
 550         }
 551
 552         for_each_cpu(cpu, drv->cpumask) {
 553                 device = &per_cpu(cpuidle_dev, cpu);
 554                 device->cpu = cpu;
 555
 556 #ifdef CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED
 557                 /*
 558                  * On multiplatform for ARM, the coupled idle states could be
 559                  * enabled in the kernel even if the cpuidle driver does not
 560                  * use it. Note, coupled_cpus is a struct copy.
 561                  */
 562                 if (coupled_cpus)
 563                         device->coupled_cpus = *coupled_cpus;
 564 #endif
 565                 ret = cpuidle_register_device(device);
 566                 if (!ret)
 567                         continue;
 568
 569                 pr_err("Failed to register cpuidle device for cpu%d\n", cpu);
 570
 571                 cpuidle_unregister(drv);
 572                 break;
 573         }
 574
 575         return ret;
 576 }
 577 EXPORT_SYMBOL_GPL(cpuidle_register);
 578
 579 #ifdef CONFIG_SMP
 580
 581 /*
 582  * This function gets called when a part of the kernel has a new latency
 583  * requirement.  This means we need to get all processors out of their C-state,
 584  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
 585  * wakes them all right up.
 586  */
 587 static int cpuidle_latency_notify(struct notifier_block *b,
 588                 unsigned long l, void *v)
 589 {
 590         wake_up_all_idle_cpus();
 591         return NOTIFY_OK;
 592 }
 593
 594 static struct notifier_block cpuidle_latency_notifier = {
 595         .notifier_call = cpuidle_latency_notify,
 596 };
 597
 598 static inline void latency_notifier_init(struct notifier_block *n)
 599 {
 600         pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n);
 601 }
 602
 603 #else /* CONFIG_SMP */
 604
 605 #define latency_notifier_init(x) do { } while (0)
 606
 607 #endif /* CONFIG_SMP */
 608
 609 /**
 610  * cpuidle_init - core initializer
 611  */
 612 static int __init cpuidle_init(void)
 613 {
 614         int ret;
 615
 616         if (cpuidle_disabled())
 617                 return -ENODEV;
 618
 619         ret = cpuidle_add_interface(cpu_subsys.dev_root);
 620         if (ret)
 621                 return ret;
 622
 623         latency_notifier_init(&cpuidle_latency_notifier);
 624
 625         return 0;
 626 }
 627
 628 module_param(off, int, 0444);
 629 core_initcall(cpuidle_init);