cpuidle.c revision 4202735e8ab6ecfb0381631a0d0b58fefe0bd4e2
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_params.h> 16#include <linux/cpu.h> 17#include <linux/cpuidle.h> 18#include <linux/ktime.h> 19#include <linux/hrtimer.h> 20#include <trace/events/power.h> 21 22#include "cpuidle.h" 23 24DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices); 25 26DEFINE_MUTEX(cpuidle_lock); 27LIST_HEAD(cpuidle_detected_devices); 28 29static int enabled_devices; 30static int off __read_mostly; 31static int initialized __read_mostly; 32 33int cpuidle_disabled(void) 34{ 35 return off; 36} 37void disable_cpuidle(void) 38{ 39 off = 1; 40} 41 42#if defined(CONFIG_ARCH_HAS_CPU_IDLE_WAIT) 43static void cpuidle_kick_cpus(void) 44{ 45 cpu_idle_wait(); 46} 47#elif defined(CONFIG_SMP) 48# error "Arch needs cpu_idle_wait() equivalent here" 49#else /* !CONFIG_ARCH_HAS_CPU_IDLE_WAIT && !CONFIG_SMP */ 50static void cpuidle_kick_cpus(void) {} 51#endif 52 53static int __cpuidle_register_device(struct cpuidle_device *dev); 54 55/** 56 * cpuidle_idle_call - the main idle loop 57 * 58 * NOTE: no locks or semaphores should be used here 59 * return non-zero on failure 60 */ 61int cpuidle_idle_call(void) 62{ 63 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices); 64 struct cpuidle_state *target_state; 65 int next_state, entered_state; 66 67 if (off) 68 return -ENODEV; 69 70 if (!initialized) 71 return -ENODEV; 72 73 /* check if the device is ready */ 74 if (!dev || !dev->enabled) 75 return -EBUSY; 76 77#if 0 78 /* shows regressions, re-enable for 2.6.29 */ 79 /* 80 * run any timers that can be run now, at this point 81 * before calculating the idle duration etc. 82 */ 83 hrtimer_peek_ahead_timers(); 84#endif 85 86 /* ask the governor for the next state */ 87 next_state = cpuidle_curr_governor->select(dev); 88 if (need_resched()) { 89 local_irq_enable(); 90 return 0; 91 } 92 93 target_state = &dev->states[next_state]; 94 95 trace_power_start(POWER_CSTATE, next_state, dev->cpu); 96 trace_cpu_idle(next_state, dev->cpu); 97 98 entered_state = target_state->enter(dev, next_state); 99 100 trace_power_end(dev->cpu); 101 trace_cpu_idle(PWR_EVENT_EXIT, dev->cpu); 102 103 if (entered_state >= 0) { 104 /* Update cpuidle counters */ 105 /* This can be moved to within driver enter routine 106 * but that results in multiple copies of same code. 107 */ 108 dev->states_usage[entered_state].time += 109 (unsigned long long)dev->last_residency; 110 dev->states_usage[entered_state].usage++; 111 } 112 113 /* give the governor an opportunity to reflect on the outcome */ 114 if (cpuidle_curr_governor->reflect) 115 cpuidle_curr_governor->reflect(dev, entered_state); 116 117 return 0; 118} 119 120/** 121 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler 122 */ 123void cpuidle_install_idle_handler(void) 124{ 125 if (enabled_devices) { 126 /* Make sure all changes finished before we switch to new idle */ 127 smp_wmb(); 128 initialized = 1; 129 } 130} 131 132/** 133 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler 134 */ 135void cpuidle_uninstall_idle_handler(void) 136{ 137 if (enabled_devices) { 138 initialized = 0; 139 cpuidle_kick_cpus(); 140 } 141} 142 143/** 144 * cpuidle_pause_and_lock - temporarily disables CPUIDLE 145 */ 146void cpuidle_pause_and_lock(void) 147{ 148 mutex_lock(&cpuidle_lock); 149 cpuidle_uninstall_idle_handler(); 150} 151 152EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock); 153 154/** 155 * cpuidle_resume_and_unlock - resumes CPUIDLE operation 156 */ 157void cpuidle_resume_and_unlock(void) 158{ 159 cpuidle_install_idle_handler(); 160 mutex_unlock(&cpuidle_lock); 161} 162 163EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock); 164 165#ifdef CONFIG_ARCH_HAS_CPU_RELAX 166static int poll_idle(struct cpuidle_device *dev, int index) 167{ 168 ktime_t t1, t2; 169 s64 diff; 170 171 t1 = ktime_get(); 172 local_irq_enable(); 173 while (!need_resched()) 174 cpu_relax(); 175 176 t2 = ktime_get(); 177 diff = ktime_to_us(ktime_sub(t2, t1)); 178 if (diff > INT_MAX) 179 diff = INT_MAX; 180 181 dev->last_residency = (int) diff; 182 183 return index; 184} 185 186static void poll_idle_init(struct cpuidle_device *dev) 187{ 188 struct cpuidle_state *state = &dev->states[0]; 189 struct cpuidle_state_usage *state_usage = &dev->states_usage[0]; 190 191 cpuidle_set_statedata(state_usage, NULL); 192 193 snprintf(state->name, CPUIDLE_NAME_LEN, "POLL"); 194 snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE"); 195 state->exit_latency = 0; 196 state->target_residency = 0; 197 state->power_usage = -1; 198 state->flags = 0; 199 state->enter = poll_idle; 200} 201#else 202static void poll_idle_init(struct cpuidle_device *dev) {} 203#endif /* CONFIG_ARCH_HAS_CPU_RELAX */ 204 205/** 206 * cpuidle_enable_device - enables idle PM for a CPU 207 * @dev: the CPU 208 * 209 * This function must be called between cpuidle_pause_and_lock and 210 * cpuidle_resume_and_unlock when used externally. 211 */ 212int cpuidle_enable_device(struct cpuidle_device *dev) 213{ 214 int ret, i; 215 216 if (dev->enabled) 217 return 0; 218 if (!cpuidle_get_driver() || !cpuidle_curr_governor) 219 return -EIO; 220 if (!dev->state_count) 221 return -EINVAL; 222 223 if (dev->registered == 0) { 224 ret = __cpuidle_register_device(dev); 225 if (ret) 226 return ret; 227 } 228 229 poll_idle_init(dev); 230 231 if ((ret = cpuidle_add_state_sysfs(dev))) 232 return ret; 233 234 if (cpuidle_curr_governor->enable && 235 (ret = cpuidle_curr_governor->enable(dev))) 236 goto fail_sysfs; 237 238 for (i = 0; i < dev->state_count; i++) { 239 dev->states_usage[i].usage = 0; 240 dev->states_usage[i].time = 0; 241 } 242 dev->last_residency = 0; 243 244 smp_wmb(); 245 246 dev->enabled = 1; 247 248 enabled_devices++; 249 return 0; 250 251fail_sysfs: 252 cpuidle_remove_state_sysfs(dev); 253 254 return ret; 255} 256 257EXPORT_SYMBOL_GPL(cpuidle_enable_device); 258 259/** 260 * cpuidle_disable_device - disables idle PM for a CPU 261 * @dev: the CPU 262 * 263 * This function must be called between cpuidle_pause_and_lock and 264 * cpuidle_resume_and_unlock when used externally. 265 */ 266void cpuidle_disable_device(struct cpuidle_device *dev) 267{ 268 if (!dev->enabled) 269 return; 270 if (!cpuidle_get_driver() || !cpuidle_curr_governor) 271 return; 272 273 dev->enabled = 0; 274 275 if (cpuidle_curr_governor->disable) 276 cpuidle_curr_governor->disable(dev); 277 278 cpuidle_remove_state_sysfs(dev); 279 enabled_devices--; 280} 281 282EXPORT_SYMBOL_GPL(cpuidle_disable_device); 283 284/** 285 * __cpuidle_register_device - internal register function called before register 286 * and enable routines 287 * @dev: the cpu 288 * 289 * cpuidle_lock mutex must be held before this is called 290 */ 291static int __cpuidle_register_device(struct cpuidle_device *dev) 292{ 293 int ret; 294 struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu); 295 struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver(); 296 297 if (!sys_dev) 298 return -EINVAL; 299 if (!try_module_get(cpuidle_driver->owner)) 300 return -EINVAL; 301 302 init_completion(&dev->kobj_unregister); 303 304 /* 305 * cpuidle driver should set the dev->power_specified bit 306 * before registering the device if the driver provides 307 * power_usage numbers. 308 * 309 * For those devices whose ->power_specified is not set, 310 * we fill in power_usage with decreasing values as the 311 * cpuidle code has an implicit assumption that state Cn 312 * uses less power than C(n-1). 313 * 314 * With CONFIG_ARCH_HAS_CPU_RELAX, C0 is already assigned 315 * an power value of -1. So we use -2, -3, etc, for other 316 * c-states. 317 */ 318 if (!dev->power_specified) { 319 int i; 320 for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++) 321 dev->states[i].power_usage = -1 - i; 322 } 323 324 per_cpu(cpuidle_devices, dev->cpu) = dev; 325 list_add(&dev->device_list, &cpuidle_detected_devices); 326 if ((ret = cpuidle_add_sysfs(sys_dev))) { 327 module_put(cpuidle_driver->owner); 328 return ret; 329 } 330 331 dev->registered = 1; 332 return 0; 333} 334 335/** 336 * cpuidle_register_device - registers a CPU's idle PM feature 337 * @dev: the cpu 338 */ 339int cpuidle_register_device(struct cpuidle_device *dev) 340{ 341 int ret; 342 343 mutex_lock(&cpuidle_lock); 344 345 if ((ret = __cpuidle_register_device(dev))) { 346 mutex_unlock(&cpuidle_lock); 347 return ret; 348 } 349 350 cpuidle_enable_device(dev); 351 cpuidle_install_idle_handler(); 352 353 mutex_unlock(&cpuidle_lock); 354 355 return 0; 356 357} 358 359EXPORT_SYMBOL_GPL(cpuidle_register_device); 360 361/** 362 * cpuidle_unregister_device - unregisters a CPU's idle PM feature 363 * @dev: the cpu 364 */ 365void cpuidle_unregister_device(struct cpuidle_device *dev) 366{ 367 struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu); 368 struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver(); 369 370 if (dev->registered == 0) 371 return; 372 373 cpuidle_pause_and_lock(); 374 375 cpuidle_disable_device(dev); 376 377 cpuidle_remove_sysfs(sys_dev); 378 list_del(&dev->device_list); 379 wait_for_completion(&dev->kobj_unregister); 380 per_cpu(cpuidle_devices, dev->cpu) = NULL; 381 382 cpuidle_resume_and_unlock(); 383 384 module_put(cpuidle_driver->owner); 385} 386 387EXPORT_SYMBOL_GPL(cpuidle_unregister_device); 388 389#ifdef CONFIG_SMP 390 391static void smp_callback(void *v) 392{ 393 /* we already woke the CPU up, nothing more to do */ 394} 395 396/* 397 * This function gets called when a part of the kernel has a new latency 398 * requirement. This means we need to get all processors out of their C-state, 399 * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that 400 * wakes them all right up. 401 */ 402static int cpuidle_latency_notify(struct notifier_block *b, 403 unsigned long l, void *v) 404{ 405 smp_call_function(smp_callback, NULL, 1); 406 return NOTIFY_OK; 407} 408 409static struct notifier_block cpuidle_latency_notifier = { 410 .notifier_call = cpuidle_latency_notify, 411}; 412 413static inline void latency_notifier_init(struct notifier_block *n) 414{ 415 pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n); 416} 417 418#else /* CONFIG_SMP */ 419 420#define latency_notifier_init(x) do { } while (0) 421 422#endif /* CONFIG_SMP */ 423 424/** 425 * cpuidle_init - core initializer 426 */ 427static int __init cpuidle_init(void) 428{ 429 int ret; 430 431 if (cpuidle_disabled()) 432 return -ENODEV; 433 434 ret = cpuidle_add_class_sysfs(&cpu_sysdev_class); 435 if (ret) 436 return ret; 437 438 latency_notifier_init(&cpuidle_latency_notifier); 439 440 return 0; 441} 442 443module_param(off, int, 0444); 444core_initcall(cpuidle_init); 445