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 <linux/mmu_context.h> 26#include <trace/events/power.h> 27 28#include "cpuidle.h" 29 30DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices); 31DEFINE_PER_CPU(struct cpuidle_device, cpuidle_dev); 32 33DEFINE_MUTEX(cpuidle_lock); 34LIST_HEAD(cpuidle_detected_devices); 35 36static int enabled_devices; 37static int off __read_mostly; 38static int initialized __read_mostly; 39 40int cpuidle_disabled(void) 41{ 42 return off; 43} 44void disable_cpuidle(void) 45{ 46 off = 1; 47} 48 49bool cpuidle_not_available(struct cpuidle_driver *drv, 50 struct cpuidle_device *dev) 51{ 52 return off || !initialized || !drv || !dev || !dev->enabled; 53} 54 55/** 56 * cpuidle_play_dead - cpu off-lining 57 * 58 * Returns in case of an error or no driver 59 */ 60int cpuidle_play_dead(void) 61{ 62 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices); 63 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev); 64 int i; 65 66 if (!drv) 67 return -ENODEV; 68 69 /* Find lowest-power state that supports long-term idle */ 70 for (i = drv->state_count - 1; i >= 0; i--) 71 if (drv->states[i].enter_dead) 72 return drv->states[i].enter_dead(dev, i); 73 74 return -ENODEV; 75} 76 77static int find_deepest_state(struct cpuidle_driver *drv, 78 struct cpuidle_device *dev, 79 u64 max_latency_ns, 80 unsigned int forbidden_flags, 81 bool s2idle) 82{ 83 u64 latency_req = 0; 84 int i, ret = 0; 85 86 for (i = 1; i < drv->state_count; i++) { 87 struct cpuidle_state *s = &drv->states[i]; 88 89 if (dev->states_usage[i].disable || 90 s->exit_latency_ns <= latency_req || 91 s->exit_latency_ns > max_latency_ns || 92 (s->flags & forbidden_flags) || 93 (s2idle && !s->enter_s2idle)) 94 continue; 95 96 latency_req = s->exit_latency_ns; 97 ret = i; 98 } 99 return ret; 100} 101 102/** 103 * cpuidle_use_deepest_state - Set/unset governor override mode. 104 * @latency_limit_ns: Idle state exit latency limit (or no override if 0). 105 * 106 * If @latency_limit_ns is nonzero, set the current CPU to use the deepest idle 107 * state with exit latency within @latency_limit_ns (override governors going 108 * forward), or do not override governors if it is zero. 109 */ 110void cpuidle_use_deepest_state(u64 latency_limit_ns) 111{ 112 struct cpuidle_device *dev; 113 114 preempt_disable(); 115 dev = cpuidle_get_device(); 116 if (dev) 117 dev->forced_idle_latency_limit_ns = latency_limit_ns; 118 preempt_enable(); 119} 120 121/** 122 * cpuidle_find_deepest_state - Find the deepest available idle state. 123 * @drv: cpuidle driver for the given CPU. 124 * @dev: cpuidle device for the given CPU. 125 * @latency_limit_ns: Idle state exit latency limit 126 * 127 * Return: the index of the deepest available idle state. 128 */ 129int cpuidle_find_deepest_state(struct cpuidle_driver *drv, 130 struct cpuidle_device *dev, 131 u64 latency_limit_ns) 132{ 133 return find_deepest_state(drv, dev, latency_limit_ns, 0, false); 134} 135 136#ifdef CONFIG_SUSPEND 137static void enter_s2idle_proper(struct cpuidle_driver *drv, 138 struct cpuidle_device *dev, int index) 139{ 140 ktime_t time_start, time_end; 141 struct cpuidle_state *target_state = &drv->states[index]; 142 143 time_start = ns_to_ktime(local_clock()); 144 145 tick_freeze(); 146 /* 147 * The state used here cannot be a "coupled" one, because the "coupled" 148 * cpuidle mechanism enables interrupts and doing that with timekeeping 149 * suspended is generally unsafe. 150 */ 151 stop_critical_timings(); 152 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) 153 rcu_idle_enter(); 154 target_state->enter_s2idle(dev, drv, index); 155 if (WARN_ON_ONCE(!irqs_disabled())) 156 local_irq_disable(); 157 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) 158 rcu_idle_exit(); 159 tick_unfreeze(); 160 start_critical_timings(); 161 162 time_end = ns_to_ktime(local_clock()); 163 164 dev->states_usage[index].s2idle_time += ktime_us_delta(time_end, time_start); 165 dev->states_usage[index].s2idle_usage++; 166} 167 168/** 169 * cpuidle_enter_s2idle - Enter an idle state suitable for suspend-to-idle. 170 * @drv: cpuidle driver for the given CPU. 171 * @dev: cpuidle device for the given CPU. 172 * 173 * If there are states with the ->enter_s2idle callback, find the deepest of 174 * them and enter it with frozen tick. 175 */ 176int cpuidle_enter_s2idle(struct cpuidle_driver *drv, struct cpuidle_device *dev) 177{ 178 int index; 179 180 /* 181 * Find the deepest state with ->enter_s2idle present, which guarantees 182 * that interrupts won't be enabled when it exits and allows the tick to 183 * be frozen safely. 184 */ 185 index = find_deepest_state(drv, dev, U64_MAX, 0, true); 186 if (index > 0) { 187 enter_s2idle_proper(drv, dev, index); 188 local_irq_enable(); 189 } 190 return index; 191} 192#endif /* CONFIG_SUSPEND */ 193 194/** 195 * cpuidle_enter_state - enter the state and update stats 196 * @dev: cpuidle device for this cpu 197 * @drv: cpuidle driver for this cpu 198 * @index: index into the states table in @drv of the state to enter 199 */ 200int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv, 201 int index) 202{ 203 int entered_state; 204 205 struct cpuidle_state *target_state = &drv->states[index]; 206 bool broadcast = !!(target_state->flags & CPUIDLE_FLAG_TIMER_STOP); 207 ktime_t time_start, time_end; 208 209 /* 210 * Tell the time framework to switch to a broadcast timer because our 211 * local timer will be shut down. If a local timer is used from another 212 * CPU as a broadcast timer, this call may fail if it is not available. 213 */ 214 if (broadcast && tick_broadcast_enter()) { 215 index = find_deepest_state(drv, dev, target_state->exit_latency_ns, 216 CPUIDLE_FLAG_TIMER_STOP, false); 217 if (index < 0) { 218 default_idle_call(); 219 return -EBUSY; 220 } 221 target_state = &drv->states[index]; 222 broadcast = false; 223 } 224 225 if (target_state->flags & CPUIDLE_FLAG_TLB_FLUSHED) 226 leave_mm(dev->cpu); 227 228 /* Take note of the planned idle state. */ 229 sched_idle_set_state(target_state); 230 231 trace_cpu_idle(index, dev->cpu); 232 time_start = ns_to_ktime(local_clock()); 233 234 stop_critical_timings(); 235 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) 236 rcu_idle_enter(); 237 entered_state = target_state->enter(dev, drv, index); 238 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) 239 rcu_idle_exit(); 240 start_critical_timings(); 241 242 sched_clock_idle_wakeup_event(); 243 time_end = ns_to_ktime(local_clock()); 244 trace_cpu_idle(PWR_EVENT_EXIT, dev->cpu); 245 246 /* The cpu is no longer idle or about to enter idle. */ 247 sched_idle_set_state(NULL); 248 249 if (broadcast) { 250 if (WARN_ON_ONCE(!irqs_disabled())) 251 local_irq_disable(); 252 253 tick_broadcast_exit(); 254 } 255 256 if (!cpuidle_state_is_coupled(drv, index)) 257 local_irq_enable(); 258 259 if (entered_state >= 0) { 260 s64 diff, delay = drv->states[entered_state].exit_latency_ns; 261 int i; 262 263 /* 264 * Update cpuidle counters 265 * This can be moved to within driver enter routine, 266 * but that results in multiple copies of same code. 267 */ 268 diff = ktime_sub(time_end, time_start); 269 270 dev->last_residency_ns = diff; 271 dev->states_usage[entered_state].time_ns += diff; 272 dev->states_usage[entered_state].usage++; 273 274 if (diff < drv->states[entered_state].target_residency_ns) { 275 for (i = entered_state - 1; i >= 0; i--) { 276 if (dev->states_usage[i].disable) 277 continue; 278 279 /* Shallower states are enabled, so update. */ 280 dev->states_usage[entered_state].above++; 281 break; 282 } 283 } else if (diff > delay) { 284 for (i = entered_state + 1; i < drv->state_count; i++) { 285 if (dev->states_usage[i].disable) 286 continue; 287 288 /* 289 * Update if a deeper state would have been a 290 * better match for the observed idle duration. 291 */ 292 if (diff - delay >= drv->states[i].target_residency_ns) 293 dev->states_usage[entered_state].below++; 294 295 break; 296 } 297 } 298 } else { 299 dev->last_residency_ns = 0; 300 dev->states_usage[index].rejected++; 301 } 302 303 return entered_state; 304} 305 306/** 307 * cpuidle_select - ask the cpuidle framework to choose an idle state 308 * 309 * @drv: the cpuidle driver 310 * @dev: the cpuidle device 311 * @stop_tick: indication on whether or not to stop the tick 312 * 313 * Returns the index of the idle state. The return value must not be negative. 314 * 315 * The memory location pointed to by @stop_tick is expected to be written the 316 * 'false' boolean value if the scheduler tick should not be stopped before 317 * entering the returned state. 318 */ 319int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, 320 bool *stop_tick) 321{ 322 return cpuidle_curr_governor->select(drv, dev, stop_tick); 323} 324 325/** 326 * cpuidle_enter - enter into the specified idle state 327 * 328 * @drv: the cpuidle driver tied with the cpu 329 * @dev: the cpuidle device 330 * @index: the index in the idle state table 331 * 332 * Returns the index in the idle state, < 0 in case of error. 333 * The error code depends on the backend driver 334 */ 335int cpuidle_enter(struct cpuidle_driver *drv, struct cpuidle_device *dev, 336 int index) 337{ 338 int ret = 0; 339 340 /* 341 * Store the next hrtimer, which becomes either next tick or the next 342 * timer event, whatever expires first. Additionally, to make this data 343 * useful for consumers outside cpuidle, we rely on that the governor's 344 * ->select() callback have decided, whether to stop the tick or not. 345 */ 346 WRITE_ONCE(dev->next_hrtimer, tick_nohz_get_next_hrtimer()); 347 348 if (cpuidle_state_is_coupled(drv, index)) 349 ret = cpuidle_enter_state_coupled(dev, drv, index); 350 else 351 ret = cpuidle_enter_state(dev, drv, index); 352 353 WRITE_ONCE(dev->next_hrtimer, 0); 354 return ret; 355} 356 357/** 358 * cpuidle_reflect - tell the underlying governor what was the state 359 * we were in 360 * 361 * @dev : the cpuidle device 362 * @index: the index in the idle state table 363 * 364 */ 365void cpuidle_reflect(struct cpuidle_device *dev, int index) 366{ 367 if (cpuidle_curr_governor->reflect && index >= 0) 368 cpuidle_curr_governor->reflect(dev, index); 369} 370 371/** 372 * cpuidle_poll_time - return amount of time to poll for, 373 * governors can override dev->poll_limit_ns if necessary 374 * 375 * @drv: the cpuidle driver tied with the cpu 376 * @dev: the cpuidle device 377 * 378 */ 379u64 cpuidle_poll_time(struct cpuidle_driver *drv, 380 struct cpuidle_device *dev) 381{ 382 int i; 383 u64 limit_ns; 384 385 if (dev->poll_limit_ns) 386 return dev->poll_limit_ns; 387 388 limit_ns = TICK_NSEC; 389 for (i = 1; i < drv->state_count; i++) { 390 if (dev->states_usage[i].disable) 391 continue; 392 393 limit_ns = drv->states[i].target_residency_ns; 394 break; 395 } 396 397 dev->poll_limit_ns = limit_ns; 398 399 return dev->poll_limit_ns; 400} 401 402/** 403 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler 404 */ 405void cpuidle_install_idle_handler(void) 406{ 407 if (enabled_devices) { 408 /* Make sure all changes finished before we switch to new idle */ 409 smp_wmb(); 410 initialized = 1; 411 } 412} 413 414/** 415 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler 416 */ 417void cpuidle_uninstall_idle_handler(void) 418{ 419 if (enabled_devices) { 420 initialized = 0; 421 wake_up_all_idle_cpus(); 422 } 423 424 /* 425 * Make sure external observers (such as the scheduler) 426 * are done looking at pointed idle states. 427 */ 428 synchronize_rcu(); 429} 430 431/** 432 * cpuidle_pause_and_lock - temporarily disables CPUIDLE 433 */ 434void cpuidle_pause_and_lock(void) 435{ 436 mutex_lock(&cpuidle_lock); 437 cpuidle_uninstall_idle_handler(); 438} 439 440EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock); 441 442/** 443 * cpuidle_resume_and_unlock - resumes CPUIDLE operation 444 */ 445void cpuidle_resume_and_unlock(void) 446{ 447 cpuidle_install_idle_handler(); 448 mutex_unlock(&cpuidle_lock); 449} 450 451EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock); 452 453/* Currently used in suspend/resume path to suspend cpuidle */ 454void cpuidle_pause(void) 455{ 456 mutex_lock(&cpuidle_lock); 457 cpuidle_uninstall_idle_handler(); 458 mutex_unlock(&cpuidle_lock); 459} 460 461/* Currently used in suspend/resume path to resume cpuidle */ 462void cpuidle_resume(void) 463{ 464 mutex_lock(&cpuidle_lock); 465 cpuidle_install_idle_handler(); 466 mutex_unlock(&cpuidle_lock); 467} 468 469/** 470 * cpuidle_enable_device - enables idle PM for a CPU 471 * @dev: the CPU 472 * 473 * This function must be called between cpuidle_pause_and_lock and 474 * cpuidle_resume_and_unlock when used externally. 475 */ 476int cpuidle_enable_device(struct cpuidle_device *dev) 477{ 478 int ret; 479 struct cpuidle_driver *drv; 480 481 if (!dev) 482 return -EINVAL; 483 484 if (dev->enabled) 485 return 0; 486 487 if (!cpuidle_curr_governor) 488 return -EIO; 489 490 drv = cpuidle_get_cpu_driver(dev); 491 492 if (!drv) 493 return -EIO; 494 495 if (!dev->registered) 496 return -EINVAL; 497 498 ret = cpuidle_add_device_sysfs(dev); 499 if (ret) 500 return ret; 501 502 if (cpuidle_curr_governor->enable) { 503 ret = cpuidle_curr_governor->enable(drv, dev); 504 if (ret) 505 goto fail_sysfs; 506 } 507 508 smp_wmb(); 509 510 dev->enabled = 1; 511 512 enabled_devices++; 513 return 0; 514 515fail_sysfs: 516 cpuidle_remove_device_sysfs(dev); 517 518 return ret; 519} 520 521EXPORT_SYMBOL_GPL(cpuidle_enable_device); 522 523/** 524 * cpuidle_disable_device - disables idle PM for a CPU 525 * @dev: the CPU 526 * 527 * This function must be called between cpuidle_pause_and_lock and 528 * cpuidle_resume_and_unlock when used externally. 529 */ 530void cpuidle_disable_device(struct cpuidle_device *dev) 531{ 532 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev); 533 534 if (!dev || !dev->enabled) 535 return; 536 537 if (!drv || !cpuidle_curr_governor) 538 return; 539 540 dev->enabled = 0; 541 542 if (cpuidle_curr_governor->disable) 543 cpuidle_curr_governor->disable(drv, dev); 544 545 cpuidle_remove_device_sysfs(dev); 546 enabled_devices--; 547} 548 549EXPORT_SYMBOL_GPL(cpuidle_disable_device); 550 551static void __cpuidle_unregister_device(struct cpuidle_device *dev) 552{ 553 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev); 554 555 list_del(&dev->device_list); 556 per_cpu(cpuidle_devices, dev->cpu) = NULL; 557 module_put(drv->owner); 558 559 dev->registered = 0; 560} 561 562static void __cpuidle_device_init(struct cpuidle_device *dev) 563{ 564 memset(dev->states_usage, 0, sizeof(dev->states_usage)); 565 dev->last_residency_ns = 0; 566 dev->next_hrtimer = 0; 567} 568 569/** 570 * __cpuidle_register_device - internal register function called before register 571 * and enable routines 572 * @dev: the cpu 573 * 574 * cpuidle_lock mutex must be held before this is called 575 */ 576static int __cpuidle_register_device(struct cpuidle_device *dev) 577{ 578 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev); 579 int i, ret; 580 581 if (!try_module_get(drv->owner)) 582 return -EINVAL; 583 584 for (i = 0; i < drv->state_count; i++) { 585 if (drv->states[i].flags & CPUIDLE_FLAG_UNUSABLE) 586 dev->states_usage[i].disable |= CPUIDLE_STATE_DISABLED_BY_DRIVER; 587 588 if (drv->states[i].flags & CPUIDLE_FLAG_OFF) 589 dev->states_usage[i].disable |= CPUIDLE_STATE_DISABLED_BY_USER; 590 } 591 592 per_cpu(cpuidle_devices, dev->cpu) = dev; 593 list_add(&dev->device_list, &cpuidle_detected_devices); 594 595 ret = cpuidle_coupled_register_device(dev); 596 if (ret) 597 __cpuidle_unregister_device(dev); 598 else 599 dev->registered = 1; 600 601 return ret; 602} 603 604/** 605 * cpuidle_register_device - registers a CPU's idle PM feature 606 * @dev: the cpu 607 */ 608int cpuidle_register_device(struct cpuidle_device *dev) 609{ 610 int ret = -EBUSY; 611 612 if (!dev) 613 return -EINVAL; 614 615 mutex_lock(&cpuidle_lock); 616 617 if (dev->registered) 618 goto out_unlock; 619 620 __cpuidle_device_init(dev); 621 622 ret = __cpuidle_register_device(dev); 623 if (ret) 624 goto out_unlock; 625 626 ret = cpuidle_add_sysfs(dev); 627 if (ret) 628 goto out_unregister; 629 630 ret = cpuidle_enable_device(dev); 631 if (ret) 632 goto out_sysfs; 633 634 cpuidle_install_idle_handler(); 635 636out_unlock: 637 mutex_unlock(&cpuidle_lock); 638 639 return ret; 640 641out_sysfs: 642 cpuidle_remove_sysfs(dev); 643out_unregister: 644 __cpuidle_unregister_device(dev); 645 goto out_unlock; 646} 647 648EXPORT_SYMBOL_GPL(cpuidle_register_device); 649 650/** 651 * cpuidle_unregister_device - unregisters a CPU's idle PM feature 652 * @dev: the cpu 653 */ 654void cpuidle_unregister_device(struct cpuidle_device *dev) 655{ 656 if (!dev || dev->registered == 0) 657 return; 658 659 cpuidle_pause_and_lock(); 660 661 cpuidle_disable_device(dev); 662 663 cpuidle_remove_sysfs(dev); 664 665 __cpuidle_unregister_device(dev); 666 667 cpuidle_coupled_unregister_device(dev); 668 669 cpuidle_resume_and_unlock(); 670} 671 672EXPORT_SYMBOL_GPL(cpuidle_unregister_device); 673 674/** 675 * cpuidle_unregister: unregister a driver and the devices. This function 676 * can be used only if the driver has been previously registered through 677 * the cpuidle_register function. 678 * 679 * @drv: a valid pointer to a struct cpuidle_driver 680 */ 681void cpuidle_unregister(struct cpuidle_driver *drv) 682{ 683 int cpu; 684 struct cpuidle_device *device; 685 686 for_each_cpu(cpu, drv->cpumask) { 687 device = &per_cpu(cpuidle_dev, cpu); 688 cpuidle_unregister_device(device); 689 } 690 691 cpuidle_unregister_driver(drv); 692} 693EXPORT_SYMBOL_GPL(cpuidle_unregister); 694 695/** 696 * cpuidle_register: registers the driver and the cpu devices with the 697 * coupled_cpus passed as parameter. This function is used for all common 698 * initialization pattern there are in the arch specific drivers. The 699 * devices is globally defined in this file. 700 * 701 * @drv : a valid pointer to a struct cpuidle_driver 702 * @coupled_cpus: a cpumask for the coupled states 703 * 704 * Returns 0 on success, < 0 otherwise 705 */ 706int cpuidle_register(struct cpuidle_driver *drv, 707 const struct cpumask *const coupled_cpus) 708{ 709 int ret, cpu; 710 struct cpuidle_device *device; 711 712 ret = cpuidle_register_driver(drv); 713 if (ret) { 714 pr_err("failed to register cpuidle driver\n"); 715 return ret; 716 } 717 718 for_each_cpu(cpu, drv->cpumask) { 719 device = &per_cpu(cpuidle_dev, cpu); 720 device->cpu = cpu; 721 722#ifdef CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED 723 /* 724 * On multiplatform for ARM, the coupled idle states could be 725 * enabled in the kernel even if the cpuidle driver does not 726 * use it. Note, coupled_cpus is a struct copy. 727 */ 728 if (coupled_cpus) 729 device->coupled_cpus = *coupled_cpus; 730#endif 731 ret = cpuidle_register_device(device); 732 if (!ret) 733 continue; 734 735 pr_err("Failed to register cpuidle device for cpu%d\n", cpu); 736 737 cpuidle_unregister(drv); 738 break; 739 } 740 741 return ret; 742} 743EXPORT_SYMBOL_GPL(cpuidle_register); 744 745/** 746 * cpuidle_init - core initializer 747 */ 748static int __init cpuidle_init(void) 749{ 750 if (cpuidle_disabled()) 751 return -ENODEV; 752 753 return cpuidle_add_interface(cpu_subsys.dev_root); 754} 755 756module_param(off, int, 0444); 757module_param_string(governor, param_governor, CPUIDLE_NAME_LEN, 0444); 758core_initcall(cpuidle_init); 759