1 /* 2 * async.c: Asynchronous function calls for boot performance 3 * 4 * (C) Copyright 2009 Intel Corporation 5 * Author: Arjan van de Ven <[email protected]> 6 * 7 * This program is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU General Public License 9 * as published by the Free Software Foundation; version 2 10 * of the License. 11 */ 12 13 14 /* 15 16 Goals and Theory of Operation 17 18 The primary goal of this feature is to reduce the kernel boot time, 19 by doing various independent hardware delays and discovery operations 20 decoupled and not strictly serialized. 21 22 More specifically, the asynchronous function call concept allows 23 certain operations (primarily during system boot) to happen 24 asynchronously, out of order, while these operations still 25 have their externally visible parts happen sequentially and in-order. 26 (not unlike how out-of-order CPUs retire their instructions in order) 27 28 Key to the asynchronous function call implementation is the concept of 29 a "sequence cookie" (which, although it has an abstracted type, can be 30 thought of as a monotonically incrementing number). 31 32 The async core will assign each scheduled event such a sequence cookie and 33 pass this to the called functions. 34 35 The asynchronously called function should before doing a globally visible 36 operation, such as registering device numbers, call the 37 async_synchronize_cookie() function and pass in its own cookie. The 38 async_synchronize_cookie() function will make sure that all asynchronous 39 operations that were scheduled prior to the operation corresponding with the 40 cookie have completed. 41 42 Subsystem/driver initialization code that scheduled asynchronous probe 43 functions, but which shares global resources with other drivers/subsystems 44 that do not use the asynchronous call feature, need to do a full 45 synchronization with the async_synchronize_full() function, before returning 46 from their init function. This is to maintain strict ordering between the 47 asynchronous and synchronous parts of the kernel. 48 49 */ 50 51 #include <linux/async.h> 52 #include <linux/atomic.h> 53 #include <linux/ktime.h> 54 #include <linux/export.h> 55 #include <linux/wait.h> 56 #include <linux/sched.h> 57 #include <linux/slab.h> 58 #include <linux/workqueue.h> 59 60 #include "workqueue_internal.h" 61 62 static async_cookie_t next_cookie = 1; 63 64 #define MAX_WORK 32768 65 66 static LIST_HEAD(async_pending); 67 static ASYNC_DOMAIN(async_running); 68 static LIST_HEAD(async_domains); 69 static DEFINE_SPINLOCK(async_lock); 70 static DEFINE_MUTEX(async_register_mutex); 71 72 struct async_entry { 73 struct list_head list; 74 struct work_struct work; 75 async_cookie_t cookie; 76 async_func_ptr *func; 77 void *data; 78 struct async_domain *running; 79 }; 80 81 static DECLARE_WAIT_QUEUE_HEAD(async_done); 82 83 static atomic_t entry_count; 84 85 86 /* 87 * MUST be called with the lock held! 88 */ 89 static async_cookie_t __lowest_in_progress(struct async_domain *running) 90 { 91 struct async_entry *entry; 92 93 if (!list_empty(&running->domain)) { 94 entry = list_first_entry(&running->domain, typeof(*entry), list); 95 return entry->cookie; 96 } 97 98 list_for_each_entry(entry, &async_pending, list) 99 if (entry->running == running) 100 return entry->cookie; 101 102 return next_cookie; /* "infinity" value */ 103 } 104 105 static async_cookie_t lowest_in_progress(struct async_domain *running) 106 { 107 unsigned long flags; 108 async_cookie_t ret; 109 110 spin_lock_irqsave(&async_lock, flags); 111 ret = __lowest_in_progress(running); 112 spin_unlock_irqrestore(&async_lock, flags); 113 return ret; 114 } 115 116 /* 117 * pick the first pending entry and run it 118 */ 119 static void async_run_entry_fn(struct work_struct *work) 120 { 121 struct async_entry *entry = 122 container_of(work, struct async_entry, work); 123 unsigned long flags; 124 ktime_t uninitialized_var(calltime), delta, rettime; 125 struct async_domain *running = entry->running; 126 127 /* 1) move self to the running queue */ 128 spin_lock_irqsave(&async_lock, flags); 129 list_move_tail(&entry->list, &running->domain); 130 spin_unlock_irqrestore(&async_lock, flags); 131 132 /* 2) run (and print duration) */ 133 if (initcall_debug && system_state == SYSTEM_BOOTING) { 134 printk(KERN_DEBUG "calling %lli_%pF @ %i\n", 135 (long long)entry->cookie, 136 entry->func, task_pid_nr(current)); 137 calltime = ktime_get(); 138 } 139 entry->func(entry->data, entry->cookie); 140 if (initcall_debug && system_state == SYSTEM_BOOTING) { 141 rettime = ktime_get(); 142 delta = ktime_sub(rettime, calltime); 143 printk(KERN_DEBUG "initcall %lli_%pF returned 0 after %lld usecs\n", 144 (long long)entry->cookie, 145 entry->func, 146 (long long)ktime_to_ns(delta) >> 10); 147 } 148 149 /* 3) remove self from the running queue */ 150 spin_lock_irqsave(&async_lock, flags); 151 list_del(&entry->list); 152 if (running->registered && --running->count == 0) 153 list_del_init(&running->node); 154 155 /* 4) free the entry */ 156 kfree(entry); 157 atomic_dec(&entry_count); 158 159 spin_unlock_irqrestore(&async_lock, flags); 160 161 /* 5) wake up any waiters */ 162 wake_up(&async_done); 163 } 164 165 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct async_domain *running) 166 { 167 struct async_entry *entry; 168 unsigned long flags; 169 async_cookie_t newcookie; 170 171 /* allow irq-off callers */ 172 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC); 173 174 /* 175 * If we're out of memory or if there's too much work 176 * pending already, we execute synchronously. 177 */ 178 if (!entry || atomic_read(&entry_count) > MAX_WORK) { 179 kfree(entry); 180 spin_lock_irqsave(&async_lock, flags); 181 newcookie = next_cookie++; 182 spin_unlock_irqrestore(&async_lock, flags); 183 184 /* low on memory.. run synchronously */ 185 ptr(data, newcookie); 186 return newcookie; 187 } 188 INIT_WORK(&entry->work, async_run_entry_fn); 189 entry->func = ptr; 190 entry->data = data; 191 entry->running = running; 192 193 spin_lock_irqsave(&async_lock, flags); 194 newcookie = entry->cookie = next_cookie++; 195 list_add_tail(&entry->list, &async_pending); 196 if (running->registered && running->count++ == 0) 197 list_add_tail(&running->node, &async_domains); 198 atomic_inc(&entry_count); 199 spin_unlock_irqrestore(&async_lock, flags); 200 201 /* schedule for execution */ 202 queue_work(system_unbound_wq, &entry->work); 203 204 return newcookie; 205 } 206 207 /** 208 * async_schedule - schedule a function for asynchronous execution 209 * @ptr: function to execute asynchronously 210 * @data: data pointer to pass to the function 211 * 212 * Returns an async_cookie_t that may be used for checkpointing later. 213 * Note: This function may be called from atomic or non-atomic contexts. 214 */ 215 async_cookie_t async_schedule(async_func_ptr *ptr, void *data) 216 { 217 return __async_schedule(ptr, data, &async_running); 218 } 219 EXPORT_SYMBOL_GPL(async_schedule); 220 221 /** 222 * async_schedule_domain - schedule a function for asynchronous execution within a certain domain 223 * @ptr: function to execute asynchronously 224 * @data: data pointer to pass to the function 225 * @running: running list for the domain 226 * 227 * Returns an async_cookie_t that may be used for checkpointing later. 228 * @running may be used in the async_synchronize_*_domain() functions 229 * to wait within a certain synchronization domain rather than globally. 230 * A synchronization domain is specified via the running queue @running to use. 231 * Note: This function may be called from atomic or non-atomic contexts. 232 */ 233 async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data, 234 struct async_domain *running) 235 { 236 return __async_schedule(ptr, data, running); 237 } 238 EXPORT_SYMBOL_GPL(async_schedule_domain); 239 240 /** 241 * async_synchronize_full - synchronize all asynchronous function calls 242 * 243 * This function waits until all asynchronous function calls have been done. 244 */ 245 void async_synchronize_full(void) 246 { 247 mutex_lock(&async_register_mutex); 248 do { 249 struct async_domain *domain = NULL; 250 251 spin_lock_irq(&async_lock); 252 if (!list_empty(&async_domains)) 253 domain = list_first_entry(&async_domains, typeof(*domain), node); 254 spin_unlock_irq(&async_lock); 255 256 async_synchronize_cookie_domain(next_cookie, domain); 257 } while (!list_empty(&async_domains)); 258 mutex_unlock(&async_register_mutex); 259 } 260 EXPORT_SYMBOL_GPL(async_synchronize_full); 261 262 /** 263 * async_unregister_domain - ensure no more anonymous waiters on this domain 264 * @domain: idle domain to flush out of any async_synchronize_full instances 265 * 266 * async_synchronize_{cookie|full}_domain() are not flushed since callers 267 * of these routines should know the lifetime of @domain 268 * 269 * Prefer ASYNC_DOMAIN_EXCLUSIVE() declarations over flushing 270 */ 271 void async_unregister_domain(struct async_domain *domain) 272 { 273 mutex_lock(&async_register_mutex); 274 spin_lock_irq(&async_lock); 275 WARN_ON(!domain->registered || !list_empty(&domain->node) || 276 !list_empty(&domain->domain)); 277 domain->registered = 0; 278 spin_unlock_irq(&async_lock); 279 mutex_unlock(&async_register_mutex); 280 } 281 EXPORT_SYMBOL_GPL(async_unregister_domain); 282 283 /** 284 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain 285 * @domain: running list to synchronize on 286 * 287 * This function waits until all asynchronous function calls for the 288 * synchronization domain specified by the running list @domain have been done. 289 */ 290 void async_synchronize_full_domain(struct async_domain *domain) 291 { 292 async_synchronize_cookie_domain(next_cookie, domain); 293 } 294 EXPORT_SYMBOL_GPL(async_synchronize_full_domain); 295 296 /** 297 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing 298 * @cookie: async_cookie_t to use as checkpoint 299 * @running: running list to synchronize on 300 * 301 * This function waits until all asynchronous function calls for the 302 * synchronization domain specified by running list @running submitted 303 * prior to @cookie have been done. 304 */ 305 void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *running) 306 { 307 ktime_t uninitialized_var(starttime), delta, endtime; 308 309 if (!running) 310 return; 311 312 if (initcall_debug && system_state == SYSTEM_BOOTING) { 313 printk(KERN_DEBUG "async_waiting @ %i\n", task_pid_nr(current)); 314 starttime = ktime_get(); 315 } 316 317 wait_event(async_done, lowest_in_progress(running) >= cookie); 318 319 if (initcall_debug && system_state == SYSTEM_BOOTING) { 320 endtime = ktime_get(); 321 delta = ktime_sub(endtime, starttime); 322 323 printk(KERN_DEBUG "async_continuing @ %i after %lli usec\n", 324 task_pid_nr(current), 325 (long long)ktime_to_ns(delta) >> 10); 326 } 327 } 328 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain); 329 330 /** 331 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing 332 * @cookie: async_cookie_t to use as checkpoint 333 * 334 * This function waits until all asynchronous function calls prior to @cookie 335 * have been done. 336 */ 337 void async_synchronize_cookie(async_cookie_t cookie) 338 { 339 async_synchronize_cookie_domain(cookie, &async_running); 340 } 341 EXPORT_SYMBOL_GPL(async_synchronize_cookie); 342 343 /** 344 * current_is_async - is %current an async worker task? 345 * 346 * Returns %true if %current is an async worker task. 347 */ 348 bool current_is_async(void) 349 { 350 struct worker *worker = current_wq_worker(); 351 352 return worker && worker->current_func == async_run_entry_fn; 353 } 354