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