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