1 /* 2 * async.c: Asynchronous function calls for boot performance 3 * 4 * (C) Copyright 2009 Intel Corporation 5 * Author: Arjan van de Ven <arjan@linux.intel.com> 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 static async_cookie_t next_cookie = 1; 61 62 #define MAX_WORK 32768 63 64 static LIST_HEAD(async_pending); 65 static LIST_HEAD(async_running); 66 static DEFINE_SPINLOCK(async_lock); 67 68 struct async_entry { 69 struct list_head list; 70 struct work_struct work; 71 async_cookie_t cookie; 72 async_func_ptr *func; 73 void *data; 74 struct list_head *running; 75 }; 76 77 static DECLARE_WAIT_QUEUE_HEAD(async_done); 78 79 static atomic_t entry_count; 80 81 82 /* 83 * MUST be called with the lock held! 84 */ 85 static async_cookie_t __lowest_in_progress(struct list_head *running) 86 { 87 struct async_entry *entry; 88 89 if (!list_empty(running)) { 90 entry = list_first_entry(running, 91 struct async_entry, list); 92 return entry->cookie; 93 } 94 95 list_for_each_entry(entry, &async_pending, list) 96 if (entry->running == running) 97 return entry->cookie; 98 99 return next_cookie; /* "infinity" value */ 100 } 101 102 static async_cookie_t lowest_in_progress(struct list_head *running) 103 { 104 unsigned long flags; 105 async_cookie_t ret; 106 107 spin_lock_irqsave(&async_lock, flags); 108 ret = __lowest_in_progress(running); 109 spin_unlock_irqrestore(&async_lock, flags); 110 return ret; 111 } 112 113 /* 114 * pick the first pending entry and run it 115 */ 116 static void async_run_entry_fn(struct work_struct *work) 117 { 118 struct async_entry *entry = 119 container_of(work, struct async_entry, work); 120 unsigned long flags; 121 ktime_t uninitialized_var(calltime), delta, rettime; 122 123 /* 1) move self to the running queue */ 124 spin_lock_irqsave(&async_lock, flags); 125 list_move_tail(&entry->list, entry->running); 126 spin_unlock_irqrestore(&async_lock, flags); 127 128 /* 2) run (and print duration) */ 129 if (initcall_debug && system_state == SYSTEM_BOOTING) { 130 printk(KERN_DEBUG "calling %lli_%pF @ %i\n", 131 (long long)entry->cookie, 132 entry->func, task_pid_nr(current)); 133 calltime = ktime_get(); 134 } 135 entry->func(entry->data, entry->cookie); 136 if (initcall_debug && system_state == SYSTEM_BOOTING) { 137 rettime = ktime_get(); 138 delta = ktime_sub(rettime, calltime); 139 printk(KERN_DEBUG "initcall %lli_%pF returned 0 after %lld usecs\n", 140 (long long)entry->cookie, 141 entry->func, 142 (long long)ktime_to_ns(delta) >> 10); 143 } 144 145 /* 3) remove self from the running queue */ 146 spin_lock_irqsave(&async_lock, flags); 147 list_del(&entry->list); 148 149 /* 4) free the entry */ 150 kfree(entry); 151 atomic_dec(&entry_count); 152 153 spin_unlock_irqrestore(&async_lock, flags); 154 155 /* 5) wake up any waiters */ 156 wake_up(&async_done); 157 } 158 159 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running) 160 { 161 struct async_entry *entry; 162 unsigned long flags; 163 async_cookie_t newcookie; 164 165 /* allow irq-off callers */ 166 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC); 167 168 /* 169 * If we're out of memory or if there's too much work 170 * pending already, we execute synchronously. 171 */ 172 if (!entry || atomic_read(&entry_count) > MAX_WORK) { 173 kfree(entry); 174 spin_lock_irqsave(&async_lock, flags); 175 newcookie = next_cookie++; 176 spin_unlock_irqrestore(&async_lock, flags); 177 178 /* low on memory.. run synchronously */ 179 ptr(data, newcookie); 180 return newcookie; 181 } 182 INIT_WORK(&entry->work, async_run_entry_fn); 183 entry->func = ptr; 184 entry->data = data; 185 entry->running = running; 186 187 spin_lock_irqsave(&async_lock, flags); 188 newcookie = entry->cookie = next_cookie++; 189 list_add_tail(&entry->list, &async_pending); 190 atomic_inc(&entry_count); 191 spin_unlock_irqrestore(&async_lock, flags); 192 193 /* schedule for execution */ 194 queue_work(system_unbound_wq, &entry->work); 195 196 return newcookie; 197 } 198 199 /** 200 * async_schedule - schedule a function for asynchronous execution 201 * @ptr: function to execute asynchronously 202 * @data: data pointer to pass to the function 203 * 204 * Returns an async_cookie_t that may be used for checkpointing later. 205 * Note: This function may be called from atomic or non-atomic contexts. 206 */ 207 async_cookie_t async_schedule(async_func_ptr *ptr, void *data) 208 { 209 return __async_schedule(ptr, data, &async_running); 210 } 211 EXPORT_SYMBOL_GPL(async_schedule); 212 213 /** 214 * async_schedule_domain - schedule a function for asynchronous execution within a certain domain 215 * @ptr: function to execute asynchronously 216 * @data: data pointer to pass to the function 217 * @running: running list for the domain 218 * 219 * Returns an async_cookie_t that may be used for checkpointing later. 220 * @running may be used in the async_synchronize_*_domain() functions 221 * to wait within a certain synchronization domain rather than globally. 222 * A synchronization domain is specified via the running queue @running to use. 223 * Note: This function may be called from atomic or non-atomic contexts. 224 */ 225 async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data, 226 struct list_head *running) 227 { 228 return __async_schedule(ptr, data, running); 229 } 230 EXPORT_SYMBOL_GPL(async_schedule_domain); 231 232 /** 233 * async_synchronize_full - synchronize all asynchronous function calls 234 * 235 * This function waits until all asynchronous function calls have been done. 236 */ 237 void async_synchronize_full(void) 238 { 239 do { 240 async_synchronize_cookie(next_cookie); 241 } while (!list_empty(&async_running) || !list_empty(&async_pending)); 242 } 243 EXPORT_SYMBOL_GPL(async_synchronize_full); 244 245 /** 246 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain 247 * @list: running list to synchronize on 248 * 249 * This function waits until all asynchronous function calls for the 250 * synchronization domain specified by the running list @list have been done. 251 */ 252 void async_synchronize_full_domain(struct list_head *list) 253 { 254 async_synchronize_cookie_domain(next_cookie, list); 255 } 256 EXPORT_SYMBOL_GPL(async_synchronize_full_domain); 257 258 /** 259 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing 260 * @cookie: async_cookie_t to use as checkpoint 261 * @running: running list to synchronize on 262 * 263 * This function waits until all asynchronous function calls for the 264 * synchronization domain specified by the running list @list submitted 265 * prior to @cookie have been done. 266 */ 267 void async_synchronize_cookie_domain(async_cookie_t cookie, 268 struct list_head *running) 269 { 270 ktime_t uninitialized_var(starttime), delta, endtime; 271 272 if (initcall_debug && system_state == SYSTEM_BOOTING) { 273 printk(KERN_DEBUG "async_waiting @ %i\n", task_pid_nr(current)); 274 starttime = ktime_get(); 275 } 276 277 wait_event(async_done, lowest_in_progress(running) >= cookie); 278 279 if (initcall_debug && system_state == SYSTEM_BOOTING) { 280 endtime = ktime_get(); 281 delta = ktime_sub(endtime, starttime); 282 283 printk(KERN_DEBUG "async_continuing @ %i after %lli usec\n", 284 task_pid_nr(current), 285 (long long)ktime_to_ns(delta) >> 10); 286 } 287 } 288 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain); 289 290 /** 291 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing 292 * @cookie: async_cookie_t to use as checkpoint 293 * 294 * This function waits until all asynchronous function calls prior to @cookie 295 * have been done. 296 */ 297 void async_synchronize_cookie(async_cookie_t cookie) 298 { 299 async_synchronize_cookie_domain(cookie, &async_running); 300 } 301 EXPORT_SYMBOL_GPL(async_synchronize_cookie); 302