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 25 DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices); 26 27 DEFINE_MUTEX(cpuidle_lock); 28 LIST_HEAD(cpuidle_detected_devices); 29 30 static int enabled_devices; 31 static int off __read_mostly; 32 static int initialized __read_mostly; 33 34 int cpuidle_disabled(void) 35 { 36 return off; 37 } 38 void disable_cpuidle(void) 39 { 40 off = 1; 41 } 42 43 #if defined(CONFIG_ARCH_HAS_CPU_IDLE_WAIT) 44 static void cpuidle_kick_cpus(void) 45 { 46 cpu_idle_wait(); 47 } 48 #elif defined(CONFIG_SMP) 49 # error "Arch needs cpu_idle_wait() equivalent here" 50 #else /* !CONFIG_ARCH_HAS_CPU_IDLE_WAIT && !CONFIG_SMP */ 51 static void cpuidle_kick_cpus(void) {} 52 #endif 53 54 static int __cpuidle_register_device(struct cpuidle_device *dev); 55 56 /** 57 * cpuidle_idle_call - the main idle loop 58 * 59 * NOTE: no locks or semaphores should be used here 60 * return non-zero on failure 61 */ 62 int cpuidle_idle_call(void) 63 { 64 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices); 65 struct cpuidle_state *target_state; 66 int next_state; 67 68 if (off) 69 return -ENODEV; 70 71 if (!initialized) 72 return -ENODEV; 73 74 /* check if the device is ready */ 75 if (!dev || !dev->enabled) 76 return -EBUSY; 77 78 #if 0 79 /* shows regressions, re-enable for 2.6.29 */ 80 /* 81 * run any timers that can be run now, at this point 82 * before calculating the idle duration etc. 83 */ 84 hrtimer_peek_ahead_timers(); 85 #endif 86 87 /* 88 * Call the device's prepare function before calling the 89 * governor's select function. ->prepare gives the device's 90 * cpuidle driver a chance to update any dynamic information 91 * of its cpuidle states for the current idle period, e.g. 92 * state availability, latencies, residencies, etc. 93 */ 94 if (dev->prepare) 95 dev->prepare(dev); 96 97 /* ask the governor for the next state */ 98 next_state = cpuidle_curr_governor->select(dev); 99 if (need_resched()) { 100 local_irq_enable(); 101 return 0; 102 } 103 104 target_state = &dev->states[next_state]; 105 106 /* enter the state and update stats */ 107 dev->last_state = target_state; 108 109 trace_power_start(POWER_CSTATE, next_state, dev->cpu); 110 trace_cpu_idle(next_state, dev->cpu); 111 112 dev->last_residency = target_state->enter(dev, target_state); 113 114 trace_power_end(dev->cpu); 115 trace_cpu_idle(PWR_EVENT_EXIT, dev->cpu); 116 117 if (dev->last_state) 118 target_state = dev->last_state; 119 120 target_state->time += (unsigned long long)dev->last_residency; 121 target_state->usage++; 122 123 /* give the governor an opportunity to reflect on the outcome */ 124 if (cpuidle_curr_governor->reflect) 125 cpuidle_curr_governor->reflect(dev); 126 127 return 0; 128 } 129 130 /** 131 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler 132 */ 133 void cpuidle_install_idle_handler(void) 134 { 135 if (enabled_devices) { 136 /* Make sure all changes finished before we switch to new idle */ 137 smp_wmb(); 138 initialized = 1; 139 } 140 } 141 142 /** 143 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler 144 */ 145 void cpuidle_uninstall_idle_handler(void) 146 { 147 if (enabled_devices) { 148 initialized = 0; 149 cpuidle_kick_cpus(); 150 } 151 } 152 153 /** 154 * cpuidle_pause_and_lock - temporarily disables CPUIDLE 155 */ 156 void cpuidle_pause_and_lock(void) 157 { 158 mutex_lock(&cpuidle_lock); 159 cpuidle_uninstall_idle_handler(); 160 } 161 162 EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock); 163 164 /** 165 * cpuidle_resume_and_unlock - resumes CPUIDLE operation 166 */ 167 void cpuidle_resume_and_unlock(void) 168 { 169 cpuidle_install_idle_handler(); 170 mutex_unlock(&cpuidle_lock); 171 } 172 173 EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock); 174 175 #ifdef CONFIG_ARCH_HAS_CPU_RELAX 176 static int poll_idle(struct cpuidle_device *dev, struct cpuidle_state *st) 177 { 178 ktime_t t1, t2; 179 s64 diff; 180 int ret; 181 182 t1 = ktime_get(); 183 local_irq_enable(); 184 while (!need_resched()) 185 cpu_relax(); 186 187 t2 = ktime_get(); 188 diff = ktime_to_us(ktime_sub(t2, t1)); 189 if (diff > INT_MAX) 190 diff = INT_MAX; 191 192 ret = (int) diff; 193 return ret; 194 } 195 196 static void poll_idle_init(struct cpuidle_device *dev) 197 { 198 struct cpuidle_state *state = &dev->states[0]; 199 200 cpuidle_set_statedata(state, NULL); 201 202 snprintf(state->name, CPUIDLE_NAME_LEN, "POLL"); 203 snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE"); 204 state->exit_latency = 0; 205 state->target_residency = 0; 206 state->power_usage = -1; 207 state->flags = 0; 208 state->enter = poll_idle; 209 } 210 #else 211 static void poll_idle_init(struct cpuidle_device *dev) {} 212 #endif /* CONFIG_ARCH_HAS_CPU_RELAX */ 213 214 /** 215 * cpuidle_enable_device - enables idle PM for a CPU 216 * @dev: the CPU 217 * 218 * This function must be called between cpuidle_pause_and_lock and 219 * cpuidle_resume_and_unlock when used externally. 220 */ 221 int cpuidle_enable_device(struct cpuidle_device *dev) 222 { 223 int ret, i; 224 225 if (dev->enabled) 226 return 0; 227 if (!cpuidle_get_driver() || !cpuidle_curr_governor) 228 return -EIO; 229 if (!dev->state_count) 230 return -EINVAL; 231 232 if (dev->registered == 0) { 233 ret = __cpuidle_register_device(dev); 234 if (ret) 235 return ret; 236 } 237 238 poll_idle_init(dev); 239 240 if ((ret = cpuidle_add_state_sysfs(dev))) 241 return ret; 242 243 if (cpuidle_curr_governor->enable && 244 (ret = cpuidle_curr_governor->enable(dev))) 245 goto fail_sysfs; 246 247 for (i = 0; i < dev->state_count; i++) { 248 dev->states[i].usage = 0; 249 dev->states[i].time = 0; 250 } 251 dev->last_residency = 0; 252 dev->last_state = NULL; 253 254 smp_wmb(); 255 256 dev->enabled = 1; 257 258 enabled_devices++; 259 return 0; 260 261 fail_sysfs: 262 cpuidle_remove_state_sysfs(dev); 263 264 return ret; 265 } 266 267 EXPORT_SYMBOL_GPL(cpuidle_enable_device); 268 269 /** 270 * cpuidle_disable_device - disables idle PM for a CPU 271 * @dev: the CPU 272 * 273 * This function must be called between cpuidle_pause_and_lock and 274 * cpuidle_resume_and_unlock when used externally. 275 */ 276 void cpuidle_disable_device(struct cpuidle_device *dev) 277 { 278 if (!dev->enabled) 279 return; 280 if (!cpuidle_get_driver() || !cpuidle_curr_governor) 281 return; 282 283 dev->enabled = 0; 284 285 if (cpuidle_curr_governor->disable) 286 cpuidle_curr_governor->disable(dev); 287 288 cpuidle_remove_state_sysfs(dev); 289 enabled_devices--; 290 } 291 292 EXPORT_SYMBOL_GPL(cpuidle_disable_device); 293 294 /** 295 * __cpuidle_register_device - internal register function called before register 296 * and enable routines 297 * @dev: the cpu 298 * 299 * cpuidle_lock mutex must be held before this is called 300 */ 301 static int __cpuidle_register_device(struct cpuidle_device *dev) 302 { 303 int ret; 304 struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu); 305 struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver(); 306 307 if (!sys_dev) 308 return -EINVAL; 309 if (!try_module_get(cpuidle_driver->owner)) 310 return -EINVAL; 311 312 init_completion(&dev->kobj_unregister); 313 314 /* 315 * cpuidle driver should set the dev->power_specified bit 316 * before registering the device if the driver provides 317 * power_usage numbers. 318 * 319 * For those devices whose ->power_specified is not set, 320 * we fill in power_usage with decreasing values as the 321 * cpuidle code has an implicit assumption that state Cn 322 * uses less power than C(n-1). 323 * 324 * With CONFIG_ARCH_HAS_CPU_RELAX, C0 is already assigned 325 * an power value of -1. So we use -2, -3, etc, for other 326 * c-states. 327 */ 328 if (!dev->power_specified) { 329 int i; 330 for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++) 331 dev->states[i].power_usage = -1 - i; 332 } 333 334 per_cpu(cpuidle_devices, dev->cpu) = dev; 335 list_add(&dev->device_list, &cpuidle_detected_devices); 336 if ((ret = cpuidle_add_sysfs(sys_dev))) { 337 module_put(cpuidle_driver->owner); 338 return ret; 339 } 340 341 dev->registered = 1; 342 return 0; 343 } 344 345 /** 346 * cpuidle_register_device - registers a CPU's idle PM feature 347 * @dev: the cpu 348 */ 349 int cpuidle_register_device(struct cpuidle_device *dev) 350 { 351 int ret; 352 353 mutex_lock(&cpuidle_lock); 354 355 if ((ret = __cpuidle_register_device(dev))) { 356 mutex_unlock(&cpuidle_lock); 357 return ret; 358 } 359 360 cpuidle_enable_device(dev); 361 cpuidle_install_idle_handler(); 362 363 mutex_unlock(&cpuidle_lock); 364 365 return 0; 366 367 } 368 369 EXPORT_SYMBOL_GPL(cpuidle_register_device); 370 371 /** 372 * cpuidle_unregister_device - unregisters a CPU's idle PM feature 373 * @dev: the cpu 374 */ 375 void cpuidle_unregister_device(struct cpuidle_device *dev) 376 { 377 struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu); 378 struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver(); 379 380 if (dev->registered == 0) 381 return; 382 383 cpuidle_pause_and_lock(); 384 385 cpuidle_disable_device(dev); 386 387 cpuidle_remove_sysfs(sys_dev); 388 list_del(&dev->device_list); 389 wait_for_completion(&dev->kobj_unregister); 390 per_cpu(cpuidle_devices, dev->cpu) = NULL; 391 392 cpuidle_resume_and_unlock(); 393 394 module_put(cpuidle_driver->owner); 395 } 396 397 EXPORT_SYMBOL_GPL(cpuidle_unregister_device); 398 399 #ifdef CONFIG_SMP 400 401 static void smp_callback(void *v) 402 { 403 /* we already woke the CPU up, nothing more to do */ 404 } 405 406 /* 407 * This function gets called when a part of the kernel has a new latency 408 * requirement. This means we need to get all processors out of their C-state, 409 * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that 410 * wakes them all right up. 411 */ 412 static int cpuidle_latency_notify(struct notifier_block *b, 413 unsigned long l, void *v) 414 { 415 smp_call_function(smp_callback, NULL, 1); 416 return NOTIFY_OK; 417 } 418 419 static struct notifier_block cpuidle_latency_notifier = { 420 .notifier_call = cpuidle_latency_notify, 421 }; 422 423 static inline void latency_notifier_init(struct notifier_block *n) 424 { 425 pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n); 426 } 427 428 #else /* CONFIG_SMP */ 429 430 #define latency_notifier_init(x) do { } while (0) 431 432 #endif /* CONFIG_SMP */ 433 434 /** 435 * cpuidle_init - core initializer 436 */ 437 static int __init cpuidle_init(void) 438 { 439 int ret; 440 441 if (cpuidle_disabled()) 442 return -ENODEV; 443 444 ret = cpuidle_add_class_sysfs(&cpu_sysdev_class); 445 if (ret) 446 return ret; 447 448 latency_notifier_init(&cpuidle_latency_notifier); 449 450 return 0; 451 } 452 453 module_param(off, int, 0444); 454 core_initcall(cpuidle_init); 455