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/sched/clock.h>
  16 #include <linux/notifier.h>
  17 #include <linux/pm_qos.h>
  18 #include <linux/cpu.h>
  19 #include <linux/cpuidle.h>
  20 #include <linux/ktime.h>
  21 #include <linux/hrtimer.h>
  22 #include <linux/module.h>
  23 #include <linux/suspend.h>
  24 #include <linux/tick.h>
  25 #include <trace/events/power.h>
  26
  27 #include "cpuidle.h"
  28
  29 DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
  30 DEFINE_PER_CPU(struct cpuidle_device, cpuidle_dev);
  31
  32 DEFINE_MUTEX(cpuidle_lock);
  33 LIST_HEAD(cpuidle_detected_devices);
  34
  35 static int enabled_devices;
  36 static int off __read_mostly;
  37 static int initialized __read_mostly;
  38
  39 int cpuidle_disabled(void)
  40 {
  41         return off;
  42 }
  43 void disable_cpuidle(void)
  44 {
  45         off = 1;
  46 }
  47
  48 bool cpuidle_not_available(struct cpuidle_driver *drv,
  49                            struct cpuidle_device *dev)
  50 {
  51         return off || !initialized || !drv || !dev || !dev->enabled;
  52 }
  53
  54 /**
  55  * cpuidle_play_dead - cpu off-lining
  56  *
  57  * Returns in case of an error or no driver
  58  */
  59 int cpuidle_play_dead(void)
  60 {
  61         struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
  62         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
  63         int i;
  64
  65         if (!drv)
  66                 return -ENODEV;
  67
  68         /* Find lowest-power state that supports long-term idle */
  69         for (i = drv->state_count - 1; i >= 0; i--)
  70                 if (drv->states[i].enter_dead)
  71                         return drv->states[i].enter_dead(dev, i);
  72
  73         return -ENODEV;
  74 }
  75
  76 static int find_deepest_state(struct cpuidle_driver *drv,
  77                               struct cpuidle_device *dev,
  78                               unsigned int max_latency,
  79                               unsigned int forbidden_flags,
  80                               bool freeze)
  81 {
  82         unsigned int latency_req = 0;
  83         int i, ret = 0;
  84
  85         for (i = 1; i < drv->state_count; i++) {
  86                 struct cpuidle_state *s = &drv->states[i];
  87                 struct cpuidle_state_usage *su = &dev->states_usage[i];
  88
  89                 if (s->disabled || su->disable || s->exit_latency <= latency_req
  90                     || s->exit_latency > max_latency
  91                     || (s->flags & forbidden_flags)
  92                     || (freeze && !s->enter_freeze))
  93                         continue;
  94
  95                 latency_req = s->exit_latency;
  96                 ret = i;
  97         }
  98         return ret;
  99 }
 100
 101 /**
 102  * cpuidle_use_deepest_state - Set/clear governor override flag.
 103  * @enable: New value of the flag.
 104  *
 105  * Set/unset the current CPU to use the deepest idle state (override governors
 106  * going forward if set).
 107  */
 108 void cpuidle_use_deepest_state(bool enable)
 109 {
 110         struct cpuidle_device *dev;
 111
 112         preempt_disable();
 113         dev = cpuidle_get_device();
 114         dev->use_deepest_state = enable;
 115         preempt_enable();
 116 }
 117
 118 /**
 119  * cpuidle_find_deepest_state - Find the deepest available idle state.
 120  * @drv: cpuidle driver for the given CPU.
 121  * @dev: cpuidle device for the given CPU.
 122  */
 123 int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
 124                                struct cpuidle_device *dev)
 125 {
 126         return find_deepest_state(drv, dev, UINT_MAX, 0, false);
 127 }
 128
 129 #ifdef CONFIG_SUSPEND
 130 static void enter_freeze_proper(struct cpuidle_driver *drv,
 131                                 struct cpuidle_device *dev, int index)
 132 {
 133         /*
 134          * trace_suspend_resume() called by tick_freeze() for the last CPU
 135          * executing it contains RCU usage regarded as invalid in the idle
 136          * context, so tell RCU about that.
 137          */
 138         RCU_NONIDLE(tick_freeze());
 139         /*
 140          * The state used here cannot be a "coupled" one, because the "coupled"
 141          * cpuidle mechanism enables interrupts and doing that with timekeeping
 142          * suspended is generally unsafe.
 143          */
 144         stop_critical_timings();
 145         drv->states[index].enter_freeze(dev, drv, index);
 146         WARN_ON(!irqs_disabled());
 147         /*
 148          * timekeeping_resume() that will be called by tick_unfreeze() for the
 149          * first CPU executing it calls functions containing RCU read-side
 150          * critical sections, so tell RCU about that.
 151          */
 152         RCU_NONIDLE(tick_unfreeze());
 153         start_critical_timings();
 154 }
 155
 156 /**
 157  * cpuidle_enter_freeze - Enter an idle state suitable for suspend-to-idle.
 158  * @drv: cpuidle driver for the given CPU.
 159  * @dev: cpuidle device for the given CPU.
 160  *
 161  * If there are states with the ->enter_freeze callback, find the deepest of
 162  * them and enter it with frozen tick.
 163  */
 164 int cpuidle_enter_freeze(struct cpuidle_driver *drv, struct cpuidle_device *dev)
 165 {
 166         int index;
 167
 168         /*
 169          * Find the deepest state with ->enter_freeze present, which guarantees
 170          * that interrupts won't be enabled when it exits and allows the tick to
 171          * be frozen safely.
 172          */
 173         index = find_deepest_state(drv, dev, UINT_MAX, 0, true);
 174         if (index > 0)
 175                 enter_freeze_proper(drv, dev, index);
 176
 177         return index;
 178 }
 179 #endif /* CONFIG_SUSPEND */
 180
 181 /**
 182  * cpuidle_enter_state - enter the state and update stats
 183  * @dev: cpuidle device for this cpu
 184  * @drv: cpuidle driver for this cpu
 185  * @index: index into the states table in @drv of the state to enter
 186  */
 187 int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
 188                         int index)
 189 {
 190         int entered_state;
 191
 192         struct cpuidle_state *target_state = &drv->states[index];
 193         bool broadcast = !!(target_state->flags & CPUIDLE_FLAG_TIMER_STOP);
 194         ktime_t time_start, time_end;
 195         s64 diff;
 196
 197         /*
 198          * Tell the time framework to switch to a broadcast timer because our
 199          * local timer will be shut down.  If a local timer is used from another
 200          * CPU as a broadcast timer, this call may fail if it is not available.
 201          */
 202         if (broadcast && tick_broadcast_enter()) {
 203                 index = find_deepest_state(drv, dev, target_state->exit_latency,
 204                                            CPUIDLE_FLAG_TIMER_STOP, false);
 205                 if (index < 0) {
 206                         default_idle_call();
 207                         return -EBUSY;
 208                 }
 209                 target_state = &drv->states[index];
 210         }
 211
 212         /* Take note of the planned idle state. */
 213         sched_idle_set_state(target_state);
 214
 215         trace_cpu_idle_rcuidle(index, dev->cpu);
 216         time_start = ns_to_ktime(local_clock());
 217
 218         stop_critical_timings();
 219         entered_state = target_state->enter(dev, drv, index);
 220         start_critical_timings();
 221
 222         time_end = ns_to_ktime(local_clock());
 223         trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
 224
 225         /* The cpu is no longer idle or about to enter idle. */
 226         sched_idle_set_state(NULL);
 227
 228         if (broadcast) {
 229                 if (WARN_ON_ONCE(!irqs_disabled()))
 230                         local_irq_disable();
 231
 232                 tick_broadcast_exit();
 233         }
 234
 235         if (!cpuidle_state_is_coupled(drv, index))
 236                 local_irq_enable();
 237
 238         diff = ktime_us_delta(time_end, time_start);
 239         if (diff > INT_MAX)
 240                 diff = INT_MAX;
 241
 242         dev->last_residency = (int) diff;
 243
 244         if (entered_state >= 0) {
 245                 /* Update cpuidle counters */
 246                 /* This can be moved to within driver enter routine
 247                  * but that results in multiple copies of same code.
 248                  */
 249                 dev->states_usage[entered_state].time += dev->last_residency;
 250                 dev->states_usage[entered_state].usage++;
 251         } else {
 252                 dev->last_residency = 0;
 253         }
 254
 255         return entered_state;
 256 }
 257
 258 /**
 259  * cpuidle_select - ask the cpuidle framework to choose an idle state
 260  *
 261  * @drv: the cpuidle driver
 262  * @dev: the cpuidle device
 263  *
 264  * Returns the index of the idle state.  The return value must not be negative.
 265  */
 266 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
 267 {
 268         return cpuidle_curr_governor->select(drv, dev);
 269 }
 270
 271 /**
 272  * cpuidle_enter - enter into the specified idle state
 273  *
 274  * @drv:   the cpuidle driver tied with the cpu
 275  * @dev:   the cpuidle device
 276  * @index: the index in the idle state table
 277  *
 278  * Returns the index in the idle state, < 0 in case of error.
 279  * The error code depends on the backend driver
 280  */
 281 int cpuidle_enter(struct cpuidle_driver *drv, struct cpuidle_device *dev,
 282                   int index)
 283 {
 284         if (cpuidle_state_is_coupled(drv, index))
 285                 return cpuidle_enter_state_coupled(dev, drv, index);
 286         return cpuidle_enter_state(dev, drv, index);
 287 }
 288
 289 /**
 290  * cpuidle_reflect - tell the underlying governor what was the state
 291  * we were in
 292  *
 293  * @dev  : the cpuidle device
 294  * @index: the index in the idle state table
 295  *
 296  */
 297 void cpuidle_reflect(struct cpuidle_device *dev, int index)
 298 {
 299         if (cpuidle_curr_governor->reflect && index >= 0)
 300                 cpuidle_curr_governor->reflect(dev, index);
 301 }
 302
 303 /**
 304  * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
 305  */
 306 void cpuidle_install_idle_handler(void)
 307 {
 308         if (enabled_devices) {
 309                 /* Make sure all changes finished before we switch to new idle */
 310                 smp_wmb();
 311                 initialized = 1;
 312         }
 313 }
 314
 315 /**
 316  * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
 317  */
 318 void cpuidle_uninstall_idle_handler(void)
 319 {
 320         if (enabled_devices) {
 321                 initialized = 0;
 322                 wake_up_all_idle_cpus();
 323         }
 324
 325         /*
 326          * Make sure external observers (such as the scheduler)
 327          * are done looking at pointed idle states.
 328          */
 329         synchronize_rcu();
 330 }
 331
 332 /**
 333  * cpuidle_pause_and_lock - temporarily disables CPUIDLE
 334  */
 335 void cpuidle_pause_and_lock(void)
 336 {
 337         mutex_lock(&cpuidle_lock);
 338         cpuidle_uninstall_idle_handler();
 339 }
 340
 341 EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
 342
 343 /**
 344  * cpuidle_resume_and_unlock - resumes CPUIDLE operation
 345  */
 346 void cpuidle_resume_and_unlock(void)
 347 {
 348         cpuidle_install_idle_handler();
 349         mutex_unlock(&cpuidle_lock);
 350 }
 351
 352 EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
 353
 354 /* Currently used in suspend/resume path to suspend cpuidle */
 355 void cpuidle_pause(void)
 356 {
 357         mutex_lock(&cpuidle_lock);
 358         cpuidle_uninstall_idle_handler();
 359         mutex_unlock(&cpuidle_lock);
 360 }
 361
 362 /* Currently used in suspend/resume path to resume cpuidle */
 363 void cpuidle_resume(void)
 364 {
 365         mutex_lock(&cpuidle_lock);
 366         cpuidle_install_idle_handler();
 367         mutex_unlock(&cpuidle_lock);
 368 }
 369
 370 /**
 371  * cpuidle_enable_device - enables idle PM for a CPU
 372  * @dev: the CPU
 373  *
 374  * This function must be called between cpuidle_pause_and_lock and
 375  * cpuidle_resume_and_unlock when used externally.
 376  */
 377 int cpuidle_enable_device(struct cpuidle_device *dev)
 378 {
 379         int ret;
 380         struct cpuidle_driver *drv;
 381
 382         if (!dev)
 383                 return -EINVAL;
 384
 385         if (dev->enabled)
 386                 return 0;
 387
 388         drv = cpuidle_get_cpu_driver(dev);
 389
 390         if (!drv || !cpuidle_curr_governor)
 391                 return -EIO;
 392
 393         if (!dev->registered)
 394                 return -EINVAL;
 395
 396         ret = cpuidle_add_device_sysfs(dev);
 397         if (ret)
 398                 return ret;
 399
 400         if (cpuidle_curr_governor->enable &&
 401             (ret = cpuidle_curr_governor->enable(drv, dev)))
 402                 goto fail_sysfs;
 403
 404         smp_wmb();
 405
 406         dev->enabled = 1;
 407
 408         enabled_devices++;
 409         return 0;
 410
 411 fail_sysfs:
 412         cpuidle_remove_device_sysfs(dev);
 413
 414         return ret;
 415 }
 416
 417 EXPORT_SYMBOL_GPL(cpuidle_enable_device);
 418
 419 /**
 420  * cpuidle_disable_device - disables idle PM for a CPU
 421  * @dev: the CPU
 422  *
 423  * This function must be called between cpuidle_pause_and_lock and
 424  * cpuidle_resume_and_unlock when used externally.
 425  */
 426 void cpuidle_disable_device(struct cpuidle_device *dev)
 427 {
 428         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 429
 430         if (!dev || !dev->enabled)
 431                 return;
 432
 433         if (!drv || !cpuidle_curr_governor)
 434                 return;
 435
 436         dev->enabled = 0;
 437
 438         if (cpuidle_curr_governor->disable)
 439                 cpuidle_curr_governor->disable(drv, dev);
 440
 441         cpuidle_remove_device_sysfs(dev);
 442         enabled_devices--;
 443 }
 444
 445 EXPORT_SYMBOL_GPL(cpuidle_disable_device);
 446
 447 static void __cpuidle_unregister_device(struct cpuidle_device *dev)
 448 {
 449         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 450
 451         list_del(&dev->device_list);
 452         per_cpu(cpuidle_devices, dev->cpu) = NULL;
 453         module_put(drv->owner);
 454
 455         dev->registered = 0;
 456 }
 457
 458 static void __cpuidle_device_init(struct cpuidle_device *dev)
 459 {
 460         memset(dev->states_usage, 0, sizeof(dev->states_usage));
 461         dev->last_residency = 0;
 462 }
 463
 464 /**
 465  * __cpuidle_register_device - internal register function called before register
 466  * and enable routines
 467  * @dev: the cpu
 468  *
 469  * cpuidle_lock mutex must be held before this is called
 470  */
 471 static int __cpuidle_register_device(struct cpuidle_device *dev)
 472 {
 473         int ret;
 474         struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 475
 476         if (!try_module_get(drv->owner))
 477                 return -EINVAL;
 478
 479         per_cpu(cpuidle_devices, dev->cpu) = dev;
 480         list_add(&dev->device_list, &cpuidle_detected_devices);
 481
 482         ret = cpuidle_coupled_register_device(dev);
 483         if (ret)
 484                 __cpuidle_unregister_device(dev);
 485         else
 486                 dev->registered = 1;
 487
 488         return ret;
 489 }
 490
 491 /**
 492  * cpuidle_register_device - registers a CPU's idle PM feature
 493  * @dev: the cpu
 494  */
 495 int cpuidle_register_device(struct cpuidle_device *dev)
 496 {
 497         int ret = -EBUSY;
 498
 499         if (!dev)
 500                 return -EINVAL;
 501
 502         mutex_lock(&cpuidle_lock);
 503
 504         if (dev->registered)
 505                 goto out_unlock;
 506
 507         __cpuidle_device_init(dev);
 508
 509         ret = __cpuidle_register_device(dev);
 510         if (ret)
 511                 goto out_unlock;
 512
 513         ret = cpuidle_add_sysfs(dev);
 514         if (ret)
 515                 goto out_unregister;
 516
 517         ret = cpuidle_enable_device(dev);
 518         if (ret)
 519                 goto out_sysfs;
 520
 521         cpuidle_install_idle_handler();
 522
 523 out_unlock:
 524         mutex_unlock(&cpuidle_lock);
 525
 526         return ret;
 527
 528 out_sysfs:
 529         cpuidle_remove_sysfs(dev);
 530 out_unregister:
 531         __cpuidle_unregister_device(dev);
 532         goto out_unlock;
 533 }
 534
 535 EXPORT_SYMBOL_GPL(cpuidle_register_device);
 536
 537 /**
 538  * cpuidle_unregister_device - unregisters a CPU's idle PM feature
 539  * @dev: the cpu
 540  */
 541 void cpuidle_unregister_device(struct cpuidle_device *dev)
 542 {
 543         if (!dev || dev->registered == 0)
 544                 return;
 545
 546         cpuidle_pause_and_lock();
 547
 548         cpuidle_disable_device(dev);
 549
 550         cpuidle_remove_sysfs(dev);
 551
 552         __cpuidle_unregister_device(dev);
 553
 554         cpuidle_coupled_unregister_device(dev);
 555
 556         cpuidle_resume_and_unlock();
 557 }
 558
 559 EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
 560
 561 /**
 562  * cpuidle_unregister: unregister a driver and the devices. This function
 563  * can be used only if the driver has been previously registered through
 564  * the cpuidle_register function.
 565  *
 566  * @drv: a valid pointer to a struct cpuidle_driver
 567  */
 568 void cpuidle_unregister(struct cpuidle_driver *drv)
 569 {
 570         int cpu;
 571         struct cpuidle_device *device;
 572
 573         for_each_cpu(cpu, drv->cpumask) {
 574                 device = &per_cpu(cpuidle_dev, cpu);
 575                 cpuidle_unregister_device(device);
 576         }
 577
 578         cpuidle_unregister_driver(drv);
 579 }
 580 EXPORT_SYMBOL_GPL(cpuidle_unregister);
 581
 582 /**
 583  * cpuidle_register: registers the driver and the cpu devices with the
 584  * coupled_cpus passed as parameter. This function is used for all common
 585  * initialization pattern there are in the arch specific drivers. The
 586  * devices is globally defined in this file.
 587  *
 588  * @drv         : a valid pointer to a struct cpuidle_driver
 589  * @coupled_cpus: a cpumask for the coupled states
 590  *
 591  * Returns 0 on success, < 0 otherwise
 592  */
 593 int cpuidle_register(struct cpuidle_driver *drv,
 594                      const struct cpumask *const coupled_cpus)
 595 {
 596         int ret, cpu;
 597         struct cpuidle_device *device;
 598
 599         ret = cpuidle_register_driver(drv);
 600         if (ret) {
 601                 pr_err("failed to register cpuidle driver\n");
 602                 return ret;
 603         }
 604
 605         for_each_cpu(cpu, drv->cpumask) {
 606                 device = &per_cpu(cpuidle_dev, cpu);
 607                 device->cpu = cpu;
 608
 609 #ifdef CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED
 610                 /*
 611                  * On multiplatform for ARM, the coupled idle states could be
 612                  * enabled in the kernel even if the cpuidle driver does not
 613                  * use it. Note, coupled_cpus is a struct copy.
 614                  */
 615                 if (coupled_cpus)
 616                         device->coupled_cpus = *coupled_cpus;
 617 #endif
 618                 ret = cpuidle_register_device(device);
 619                 if (!ret)
 620                         continue;
 621
 622                 pr_err("Failed to register cpuidle device for cpu%d\n", cpu);
 623
 624                 cpuidle_unregister(drv);
 625                 break;
 626         }
 627
 628         return ret;
 629 }
 630 EXPORT_SYMBOL_GPL(cpuidle_register);
 631
 632 #ifdef CONFIG_SMP
 633
 634 /*
 635  * This function gets called when a part of the kernel has a new latency
 636  * requirement.  This means we need to get all processors out of their C-state,
 637  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
 638  * wakes them all right up.
 639  */
 640 static int cpuidle_latency_notify(struct notifier_block *b,
 641                 unsigned long l, void *v)
 642 {
 643         wake_up_all_idle_cpus();
 644         return NOTIFY_OK;
 645 }
 646
 647 static struct notifier_block cpuidle_latency_notifier = {
 648         .notifier_call = cpuidle_latency_notify,
 649 };
 650
 651 static inline void latency_notifier_init(struct notifier_block *n)
 652 {
 653         pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n);
 654 }
 655
 656 #else /* CONFIG_SMP */
 657
 658 #define latency_notifier_init(x) do { } while (0)
 659
 660 #endif /* CONFIG_SMP */
 661
 662 /**
 663  * cpuidle_init - core initializer
 664  */
 665 static int __init cpuidle_init(void)
 666 {
 667         int ret;
 668
 669         if (cpuidle_disabled())
 670                 return -ENODEV;
 671
 672         ret = cpuidle_add_interface(cpu_subsys.dev_root);
 673         if (ret)
 674                 return ret;
 675
 676         latency_notifier_init(&cpuidle_latency_notifier);
 677
 678         return 0;
 679 }
 680
 681 module_param(off, int, 0444);
 682 core_initcall(cpuidle_init);