2 * async.c: Asynchronous function calls for boot performance
4 * (C) Copyright 2009 Intel Corporation
5 * Author: Arjan van de Ven <arjan@linux.intel.com>
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
16 Goals and Theory of Operation
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.
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)
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).
32 The async core will assign each scheduled event such a sequence cookie and
33 pass this to the called functions.
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.
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.
51 #include <linux/async.h>
52 #include <linux/bug.h>
53 #include <linux/module.h>
54 #include <linux/wait.h>
55 #include <linux/sched.h>
56 #include <linux/init.h>
57 #include <linux/kthread.h>
58 #include <linux/delay.h>
59 #include <asm/atomic.h>
61 static async_cookie_t next_cookie = 1;
63 #define MAX_THREADS 256
64 #define MAX_WORK 32768
66 static LIST_HEAD(async_pending);
67 static LIST_HEAD(async_running);
68 static DEFINE_SPINLOCK(async_lock);
70 static int async_enabled = 0;
73 struct list_head list;
74 async_cookie_t cookie;
77 struct list_head *running;
80 static DECLARE_WAIT_QUEUE_HEAD(async_done);
81 static DECLARE_WAIT_QUEUE_HEAD(async_new);
83 static atomic_t entry_count;
84 static atomic_t thread_count;
86 extern int initcall_debug;
90 * MUST be called with the lock held!
92 static async_cookie_t __lowest_in_progress(struct list_head *running)
94 struct async_entry *entry;
95 async_cookie_t ret = next_cookie; /* begin with "infinity" value */
97 if (!list_empty(running)) {
98 entry = list_first_entry(running,
99 struct async_entry, list);
103 if (!list_empty(&async_pending)) {
104 list_for_each_entry(entry, &async_pending, list)
105 if (entry->running == running) {
114 static async_cookie_t lowest_in_progress(struct list_head *running)
119 spin_lock_irqsave(&async_lock, flags);
120 ret = __lowest_in_progress(running);
121 spin_unlock_irqrestore(&async_lock, flags);
125 * pick the first pending entry and run it
127 static void run_one_entry(void)
130 struct async_entry *entry;
131 ktime_t calltime, delta, rettime;
133 /* 1) pick one task from the pending queue */
135 spin_lock_irqsave(&async_lock, flags);
136 if (list_empty(&async_pending))
138 entry = list_first_entry(&async_pending, struct async_entry, list);
140 /* 2) move it to the running queue */
141 list_move_tail(&entry->list, entry->running);
142 spin_unlock_irqrestore(&async_lock, flags);
144 /* 3) run it (and print duration)*/
145 if (initcall_debug && system_state == SYSTEM_BOOTING) {
146 printk("calling %lli_%pF @ %i\n", (long long)entry->cookie,
147 entry->func, task_pid_nr(current));
148 calltime = ktime_get();
150 entry->func(entry->data, entry->cookie);
151 if (initcall_debug && system_state == SYSTEM_BOOTING) {
152 rettime = ktime_get();
153 delta = ktime_sub(rettime, calltime);
154 printk("initcall %lli_%pF returned 0 after %lld usecs\n",
155 (long long)entry->cookie,
157 (long long)ktime_to_ns(delta) >> 10);
160 /* 4) remove it from the running queue */
161 spin_lock_irqsave(&async_lock, flags);
162 list_del(&entry->list);
164 /* 5) free the entry */
166 atomic_dec(&entry_count);
168 spin_unlock_irqrestore(&async_lock, flags);
170 /* 6) wake up any waiters. */
171 wake_up(&async_done);
175 spin_unlock_irqrestore(&async_lock, flags);
179 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running)
181 struct async_entry *entry;
183 async_cookie_t newcookie;
186 /* allow irq-off callers */
187 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
190 * If we're out of memory or if there's too much work
191 * pending already, we execute synchronously.
193 if (!async_enabled || !entry || atomic_read(&entry_count) > MAX_WORK) {
195 spin_lock_irqsave(&async_lock, flags);
196 newcookie = next_cookie++;
197 spin_unlock_irqrestore(&async_lock, flags);
199 /* low on memory.. run synchronously */
200 ptr(data, newcookie);
205 entry->running = running;
207 spin_lock_irqsave(&async_lock, flags);
208 newcookie = entry->cookie = next_cookie++;
209 list_add_tail(&entry->list, &async_pending);
210 atomic_inc(&entry_count);
211 spin_unlock_irqrestore(&async_lock, flags);
217 * async_schedule - schedule a function for asynchronous execution
218 * @ptr: function to execute asynchronously
219 * @data: data pointer to pass to the function
221 * Returns an async_cookie_t that may be used for checkpointing later.
222 * Note: This function may be called from atomic or non-atomic contexts.
224 async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
226 return __async_schedule(ptr, data, &async_running);
228 EXPORT_SYMBOL_GPL(async_schedule);
231 * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
232 * @ptr: function to execute asynchronously
233 * @data: data pointer to pass to the function
234 * @running: running list for the domain
236 * Returns an async_cookie_t that may be used for checkpointing later.
237 * @running may be used in the async_synchronize_*_domain() functions
238 * to wait within a certain synchronization domain rather than globally.
239 * A synchronization domain is specified via the running queue @running to use.
240 * Note: This function may be called from atomic or non-atomic contexts.
242 async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data,
243 struct list_head *running)
245 return __async_schedule(ptr, data, running);
247 EXPORT_SYMBOL_GPL(async_schedule_domain);
250 * async_synchronize_full - synchronize all asynchronous function calls
252 * This function waits until all asynchronous function calls have been done.
254 void async_synchronize_full(void)
257 async_synchronize_cookie(next_cookie);
258 } while (!list_empty(&async_running) || !list_empty(&async_pending));
260 EXPORT_SYMBOL_GPL(async_synchronize_full);
263 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
264 * @list: running list to synchronize on
266 * This function waits until all asynchronous function calls for the
267 * synchronization domain specified by the running list @list have been done.
269 void async_synchronize_full_domain(struct list_head *list)
271 async_synchronize_cookie_domain(next_cookie, list);
273 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
276 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
277 * @cookie: async_cookie_t to use as checkpoint
278 * @running: running list to synchronize on
280 * This function waits until all asynchronous function calls for the
281 * synchronization domain specified by the running list @list submitted
282 * prior to @cookie have been done.
284 void async_synchronize_cookie_domain(async_cookie_t cookie,
285 struct list_head *running)
287 ktime_t starttime, delta, endtime;
289 if (initcall_debug && system_state == SYSTEM_BOOTING) {
290 printk("async_waiting @ %i\n", task_pid_nr(current));
291 starttime = ktime_get();
294 wait_event(async_done, lowest_in_progress(running) >= cookie);
296 if (initcall_debug && system_state == SYSTEM_BOOTING) {
297 endtime = ktime_get();
298 delta = ktime_sub(endtime, starttime);
300 printk("async_continuing @ %i after %lli usec\n",
301 task_pid_nr(current),
302 (long long)ktime_to_ns(delta) >> 10);
305 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
308 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
309 * @cookie: async_cookie_t to use as checkpoint
311 * This function waits until all asynchronous function calls prior to @cookie
314 void async_synchronize_cookie(async_cookie_t cookie)
316 async_synchronize_cookie_domain(cookie, &async_running);
318 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
321 static int async_thread(void *unused)
323 DECLARE_WAITQUEUE(wq, current);
324 add_wait_queue(&async_new, &wq);
326 while (!kthread_should_stop()) {
328 set_current_state(TASK_INTERRUPTIBLE);
330 * check the list head without lock.. false positives
331 * are dealt with inside run_one_entry() while holding
335 if (!list_empty(&async_pending))
338 ret = schedule_timeout(HZ);
342 * we timed out, this means we as thread are redundant.
343 * we sign off and die, but we to avoid any races there
344 * is a last-straw check to see if work snuck in.
346 atomic_dec(&thread_count);
347 wmb(); /* manager must see our departure first */
348 if (list_empty(&async_pending))
351 * woops work came in between us timing out and us
352 * signing off; we need to stay alive and keep working.
354 atomic_inc(&thread_count);
357 remove_wait_queue(&async_new, &wq);
362 static int async_manager_thread(void *unused)
364 DECLARE_WAITQUEUE(wq, current);
365 add_wait_queue(&async_new, &wq);
367 while (!kthread_should_stop()) {
370 set_current_state(TASK_INTERRUPTIBLE);
372 tc = atomic_read(&thread_count);
374 ec = atomic_read(&entry_count);
376 while (tc < ec && tc < MAX_THREADS) {
377 if (IS_ERR(kthread_run(async_thread, NULL, "async/%i",
382 atomic_inc(&thread_count);
388 remove_wait_queue(&async_new, &wq);
393 static int __init async_init(void)
396 !IS_ERR(kthread_run(async_manager_thread, NULL, "async/mgr"));
398 WARN_ON(!async_enabled);
402 core_initcall(async_init);