cpuidle.c revision 1aef40e288acfb3cc28ff77528b34ef66683bed6
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/kernel.h> 12#include <linux/mutex.h> 13#include <linux/sched.h> 14#include <linux/notifier.h> 15#include <linux/pm_qos.h> 16#include <linux/cpu.h> 17#include <linux/cpuidle.h> 18#include <linux/ktime.h> 19#include <linux/hrtimer.h> 20#include <linux/module.h> 21#include <trace/events/power.h> 22 23#include "cpuidle.h" 24 25DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices); 26 27DEFINE_MUTEX(cpuidle_lock); 28LIST_HEAD(cpuidle_detected_devices); 29 30static int enabled_devices; 31static int off __read_mostly; 32static int initialized __read_mostly; 33 34int cpuidle_disabled(void) 35{ 36 return off; 37} 38void disable_cpuidle(void) 39{ 40 off = 1; 41} 42 43static int __cpuidle_register_device(struct cpuidle_device *dev); 44 45static inline int cpuidle_enter(struct cpuidle_device *dev, 46 struct cpuidle_driver *drv, int index) 47{ 48 struct cpuidle_state *target_state = &drv->states[index]; 49 return target_state->enter(dev, drv, index); 50} 51 52static inline int cpuidle_enter_tk(struct cpuidle_device *dev, 53 struct cpuidle_driver *drv, int index) 54{ 55 return cpuidle_wrap_enter(dev, drv, index, cpuidle_enter); 56} 57 58typedef int (*cpuidle_enter_t)(struct cpuidle_device *dev, 59 struct cpuidle_driver *drv, int index); 60 61static cpuidle_enter_t cpuidle_enter_ops; 62 63/** 64 * cpuidle_play_dead - cpu off-lining 65 * 66 * Returns in case of an error or no driver 67 */ 68int cpuidle_play_dead(void) 69{ 70 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices); 71 struct cpuidle_driver *drv = cpuidle_get_driver(); 72 int i, dead_state = -1; 73 int power_usage = -1; 74 75 if (!drv) 76 return -ENODEV; 77 78 /* Find lowest-power state that supports long-term idle */ 79 for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) { 80 struct cpuidle_state *s = &drv->states[i]; 81 82 if (s->power_usage < power_usage && s->enter_dead) { 83 power_usage = s->power_usage; 84 dead_state = i; 85 } 86 } 87 88 if (dead_state != -1) 89 return drv->states[dead_state].enter_dead(dev, dead_state); 90 91 return -ENODEV; 92} 93 94/** 95 * cpuidle_enter_state - enter the state and update stats 96 * @dev: cpuidle device for this cpu 97 * @drv: cpuidle driver for this cpu 98 * @next_state: index into drv->states of the state to enter 99 */ 100int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv, 101 int next_state) 102{ 103 int entered_state; 104 105 entered_state = cpuidle_enter_ops(dev, drv, next_state); 106 107 if (entered_state >= 0) { 108 /* Update cpuidle counters */ 109 /* This can be moved to within driver enter routine 110 * but that results in multiple copies of same code. 111 */ 112 dev->states_usage[entered_state].time += 113 (unsigned long long)dev->last_residency; 114 dev->states_usage[entered_state].usage++; 115 } else { 116 dev->last_residency = 0; 117 } 118 119 return entered_state; 120} 121 122/** 123 * cpuidle_idle_call - the main idle loop 124 * 125 * NOTE: no locks or semaphores should be used here 126 * return non-zero on failure 127 */ 128int cpuidle_idle_call(void) 129{ 130 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices); 131 struct cpuidle_driver *drv = cpuidle_get_driver(); 132 int next_state, entered_state; 133 134 if (off) 135 return -ENODEV; 136 137 if (!initialized) 138 return -ENODEV; 139 140 /* check if the device is ready */ 141 if (!dev || !dev->enabled) 142 return -EBUSY; 143 144 /* ask the governor for the next state */ 145 next_state = cpuidle_curr_governor->select(drv, dev); 146 if (need_resched()) { 147 local_irq_enable(); 148 return 0; 149 } 150 151 trace_power_start_rcuidle(POWER_CSTATE, next_state, dev->cpu); 152 trace_cpu_idle_rcuidle(next_state, dev->cpu); 153 154 if (cpuidle_state_is_coupled(dev, drv, next_state)) 155 entered_state = cpuidle_enter_state_coupled(dev, drv, 156 next_state); 157 else 158 entered_state = cpuidle_enter_state(dev, drv, next_state); 159 160 trace_power_end_rcuidle(dev->cpu); 161 trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu); 162 163 /* give the governor an opportunity to reflect on the outcome */ 164 if (cpuidle_curr_governor->reflect) 165 cpuidle_curr_governor->reflect(dev, entered_state); 166 167 return 0; 168} 169 170/** 171 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler 172 */ 173void cpuidle_install_idle_handler(void) 174{ 175 if (enabled_devices) { 176 /* Make sure all changes finished before we switch to new idle */ 177 smp_wmb(); 178 initialized = 1; 179 } 180} 181 182/** 183 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler 184 */ 185void cpuidle_uninstall_idle_handler(void) 186{ 187 if (enabled_devices) { 188 initialized = 0; 189 kick_all_cpus_sync(); 190 } 191} 192 193/** 194 * cpuidle_pause_and_lock - temporarily disables CPUIDLE 195 */ 196void cpuidle_pause_and_lock(void) 197{ 198 mutex_lock(&cpuidle_lock); 199 cpuidle_uninstall_idle_handler(); 200} 201 202EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock); 203 204/** 205 * cpuidle_resume_and_unlock - resumes CPUIDLE operation 206 */ 207void cpuidle_resume_and_unlock(void) 208{ 209 cpuidle_install_idle_handler(); 210 mutex_unlock(&cpuidle_lock); 211} 212 213EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock); 214 215/* Currently used in suspend/resume path to suspend cpuidle */ 216void cpuidle_pause(void) 217{ 218 mutex_lock(&cpuidle_lock); 219 cpuidle_uninstall_idle_handler(); 220 mutex_unlock(&cpuidle_lock); 221} 222 223/* Currently used in suspend/resume path to resume cpuidle */ 224void cpuidle_resume(void) 225{ 226 mutex_lock(&cpuidle_lock); 227 cpuidle_install_idle_handler(); 228 mutex_unlock(&cpuidle_lock); 229} 230 231/** 232 * cpuidle_wrap_enter - performs timekeeping and irqen around enter function 233 * @dev: pointer to a valid cpuidle_device object 234 * @drv: pointer to a valid cpuidle_driver object 235 * @index: index of the target cpuidle state. 236 */ 237int cpuidle_wrap_enter(struct cpuidle_device *dev, 238 struct cpuidle_driver *drv, int index, 239 int (*enter)(struct cpuidle_device *dev, 240 struct cpuidle_driver *drv, int index)) 241{ 242 ktime_t time_start, time_end; 243 s64 diff; 244 245 time_start = ktime_get(); 246 247 index = enter(dev, drv, index); 248 249 time_end = ktime_get(); 250 251 local_irq_enable(); 252 253 diff = ktime_to_us(ktime_sub(time_end, time_start)); 254 if (diff > INT_MAX) 255 diff = INT_MAX; 256 257 dev->last_residency = (int) diff; 258 259 return index; 260} 261 262#ifdef CONFIG_ARCH_HAS_CPU_RELAX 263static int poll_idle(struct cpuidle_device *dev, 264 struct cpuidle_driver *drv, int index) 265{ 266 ktime_t t1, t2; 267 s64 diff; 268 269 t1 = ktime_get(); 270 local_irq_enable(); 271 while (!need_resched()) 272 cpu_relax(); 273 274 t2 = ktime_get(); 275 diff = ktime_to_us(ktime_sub(t2, t1)); 276 if (diff > INT_MAX) 277 diff = INT_MAX; 278 279 dev->last_residency = (int) diff; 280 281 return index; 282} 283 284static void poll_idle_init(struct cpuidle_driver *drv) 285{ 286 struct cpuidle_state *state = &drv->states[0]; 287 288 snprintf(state->name, CPUIDLE_NAME_LEN, "POLL"); 289 snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE"); 290 state->exit_latency = 0; 291 state->target_residency = 0; 292 state->power_usage = -1; 293 state->flags = 0; 294 state->enter = poll_idle; 295 state->disabled = false; 296} 297#else 298static void poll_idle_init(struct cpuidle_driver *drv) {} 299#endif /* CONFIG_ARCH_HAS_CPU_RELAX */ 300 301/** 302 * cpuidle_enable_device - enables idle PM for a CPU 303 * @dev: the CPU 304 * 305 * This function must be called between cpuidle_pause_and_lock and 306 * cpuidle_resume_and_unlock when used externally. 307 */ 308int cpuidle_enable_device(struct cpuidle_device *dev) 309{ 310 int ret, i; 311 struct cpuidle_driver *drv = cpuidle_get_driver(); 312 313 if (!dev) 314 return -EINVAL; 315 316 if (dev->enabled) 317 return 0; 318 if (!drv || !cpuidle_curr_governor) 319 return -EIO; 320 if (!dev->state_count) 321 dev->state_count = drv->state_count; 322 323 if (dev->registered == 0) { 324 ret = __cpuidle_register_device(dev); 325 if (ret) 326 return ret; 327 } 328 329 cpuidle_enter_ops = drv->en_core_tk_irqen ? 330 cpuidle_enter_tk : cpuidle_enter; 331 332 poll_idle_init(drv); 333 334 if ((ret = cpuidle_add_state_sysfs(dev))) 335 return ret; 336 337 if (cpuidle_curr_governor->enable && 338 (ret = cpuidle_curr_governor->enable(drv, dev))) 339 goto fail_sysfs; 340 341 for (i = 0; i < dev->state_count; i++) { 342 dev->states_usage[i].usage = 0; 343 dev->states_usage[i].time = 0; 344 } 345 dev->last_residency = 0; 346 347 smp_wmb(); 348 349 dev->enabled = 1; 350 351 enabled_devices++; 352 return 0; 353 354fail_sysfs: 355 cpuidle_remove_state_sysfs(dev); 356 357 return ret; 358} 359 360EXPORT_SYMBOL_GPL(cpuidle_enable_device); 361 362/** 363 * cpuidle_disable_device - disables idle PM for a CPU 364 * @dev: the CPU 365 * 366 * This function must be called between cpuidle_pause_and_lock and 367 * cpuidle_resume_and_unlock when used externally. 368 */ 369void cpuidle_disable_device(struct cpuidle_device *dev) 370{ 371 if (!dev || !dev->enabled) 372 return; 373 if (!cpuidle_get_driver() || !cpuidle_curr_governor) 374 return; 375 376 dev->enabled = 0; 377 378 if (cpuidle_curr_governor->disable) 379 cpuidle_curr_governor->disable(cpuidle_get_driver(), dev); 380 381 cpuidle_remove_state_sysfs(dev); 382 enabled_devices--; 383} 384 385EXPORT_SYMBOL_GPL(cpuidle_disable_device); 386 387/** 388 * __cpuidle_register_device - internal register function called before register 389 * and enable routines 390 * @dev: the cpu 391 * 392 * cpuidle_lock mutex must be held before this is called 393 */ 394static int __cpuidle_register_device(struct cpuidle_device *dev) 395{ 396 int ret; 397 struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver(); 398 399 if (!try_module_get(cpuidle_driver->owner)) 400 return -EINVAL; 401 402 init_completion(&dev->kobj_unregister); 403 404 per_cpu(cpuidle_devices, dev->cpu) = dev; 405 list_add(&dev->device_list, &cpuidle_detected_devices); 406 ret = cpuidle_add_sysfs(dev); 407 if (ret) 408 goto err_sysfs; 409 410 ret = cpuidle_coupled_register_device(dev); 411 if (ret) 412 goto err_coupled; 413 414 dev->registered = 1; 415 return 0; 416 417err_coupled: 418 cpuidle_remove_sysfs(dev); 419 wait_for_completion(&dev->kobj_unregister); 420err_sysfs: 421 list_del(&dev->device_list); 422 per_cpu(cpuidle_devices, dev->cpu) = NULL; 423 module_put(cpuidle_driver->owner); 424 return ret; 425} 426 427/** 428 * cpuidle_register_device - registers a CPU's idle PM feature 429 * @dev: the cpu 430 */ 431int cpuidle_register_device(struct cpuidle_device *dev) 432{ 433 int ret; 434 435 if (!dev) 436 return -EINVAL; 437 438 mutex_lock(&cpuidle_lock); 439 440 if ((ret = __cpuidle_register_device(dev))) { 441 mutex_unlock(&cpuidle_lock); 442 return ret; 443 } 444 445 cpuidle_enable_device(dev); 446 cpuidle_install_idle_handler(); 447 448 mutex_unlock(&cpuidle_lock); 449 450 return 0; 451 452} 453 454EXPORT_SYMBOL_GPL(cpuidle_register_device); 455 456/** 457 * cpuidle_unregister_device - unregisters a CPU's idle PM feature 458 * @dev: the cpu 459 */ 460void cpuidle_unregister_device(struct cpuidle_device *dev) 461{ 462 struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver(); 463 464 if (dev->registered == 0) 465 return; 466 467 cpuidle_pause_and_lock(); 468 469 cpuidle_disable_device(dev); 470 471 cpuidle_remove_sysfs(dev); 472 list_del(&dev->device_list); 473 wait_for_completion(&dev->kobj_unregister); 474 per_cpu(cpuidle_devices, dev->cpu) = NULL; 475 476 cpuidle_coupled_unregister_device(dev); 477 478 cpuidle_resume_and_unlock(); 479 480 module_put(cpuidle_driver->owner); 481} 482 483EXPORT_SYMBOL_GPL(cpuidle_unregister_device); 484 485#ifdef CONFIG_SMP 486 487static void smp_callback(void *v) 488{ 489 /* we already woke the CPU up, nothing more to do */ 490} 491 492/* 493 * This function gets called when a part of the kernel has a new latency 494 * requirement. This means we need to get all processors out of their C-state, 495 * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that 496 * wakes them all right up. 497 */ 498static int cpuidle_latency_notify(struct notifier_block *b, 499 unsigned long l, void *v) 500{ 501 smp_call_function(smp_callback, NULL, 1); 502 return NOTIFY_OK; 503} 504 505static struct notifier_block cpuidle_latency_notifier = { 506 .notifier_call = cpuidle_latency_notify, 507}; 508 509static inline void latency_notifier_init(struct notifier_block *n) 510{ 511 pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n); 512} 513 514#else /* CONFIG_SMP */ 515 516#define latency_notifier_init(x) do { } while (0) 517 518#endif /* CONFIG_SMP */ 519 520/** 521 * cpuidle_init - core initializer 522 */ 523static int __init cpuidle_init(void) 524{ 525 int ret; 526 527 if (cpuidle_disabled()) 528 return -ENODEV; 529 530 ret = cpuidle_add_interface(cpu_subsys.dev_root); 531 if (ret) 532 return ret; 533 534 latency_notifier_init(&cpuidle_latency_notifier); 535 536 return 0; 537} 538 539module_param(off, int, 0444); 540core_initcall(cpuidle_init); 541