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/module.h>
53 #include <linux/wait.h>
54 #include <linux/sched.h>
55 #include <linux/init.h>
56 #include <linux/kthread.h>
57 #include <asm/atomic.h>
59 static async_cookie_t next_cookie = 1;
61 #define MAX_THREADS 256
62 #define MAX_WORK 32768
64 static LIST_HEAD(async_pending);
65 static LIST_HEAD(async_running);
66 static DEFINE_SPINLOCK(async_lock);
68 static int async_enabled = 0;
71 struct list_head list;
72 async_cookie_t cookie;
75 struct list_head *running;
78 static DECLARE_WAIT_QUEUE_HEAD(async_done);
79 static DECLARE_WAIT_QUEUE_HEAD(async_new);
81 static atomic_t entry_count;
82 static atomic_t thread_count;
84 extern int initcall_debug;
88 * MUST be called with the lock held!
90 static async_cookie_t __lowest_in_progress(struct list_head *running)
92 struct async_entry *entry;
93 if (!list_empty(running)) {
94 entry = list_first_entry(running,
95 struct async_entry, list);
97 } else if (!list_empty(&async_pending)) {
98 entry = list_first_entry(&async_pending,
99 struct async_entry, list);
100 return entry->cookie;
102 /* nothing in progress... next_cookie is "infinity" */
108 static async_cookie_t lowest_in_progress(struct list_head *running)
113 spin_lock_irqsave(&async_lock, flags);
114 ret = __lowest_in_progress(running);
115 spin_unlock_irqrestore(&async_lock, flags);
119 * pick the first pending entry and run it
121 static void run_one_entry(void)
124 struct async_entry *entry;
125 ktime_t calltime, delta, rettime;
127 /* 1) pick one task from the pending queue */
129 spin_lock_irqsave(&async_lock, flags);
130 if (list_empty(&async_pending))
132 entry = list_first_entry(&async_pending, struct async_entry, list);
134 /* 2) move it to the running queue */
135 list_del(&entry->list);
136 list_add_tail(&entry->list, &async_running);
137 spin_unlock_irqrestore(&async_lock, flags);
139 /* 3) run it (and print duration)*/
140 if (initcall_debug && system_state == SYSTEM_BOOTING) {
141 printk("calling %lli_%pF @ %i\n", entry->cookie, entry->func, task_pid_nr(current));
142 calltime = ktime_get();
144 entry->func(entry->data, entry->cookie);
145 if (initcall_debug && system_state == SYSTEM_BOOTING) {
146 rettime = ktime_get();
147 delta = ktime_sub(rettime, calltime);
148 printk("initcall %lli_%pF returned 0 after %lld usecs\n", entry->cookie,
149 entry->func, ktime_to_ns(delta) >> 10);
152 /* 4) remove it from the running queue */
153 spin_lock_irqsave(&async_lock, flags);
154 list_del(&entry->list);
156 /* 5) free the entry */
158 atomic_dec(&entry_count);
160 spin_unlock_irqrestore(&async_lock, flags);
162 /* 6) wake up any waiters. */
163 wake_up(&async_done);
167 spin_unlock_irqrestore(&async_lock, flags);
171 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running)
173 struct async_entry *entry;
175 async_cookie_t newcookie;
178 /* allow irq-off callers */
179 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
182 * If we're out of memory or if there's too much work
183 * pending already, we execute synchronously.
185 if (!async_enabled || !entry || atomic_read(&entry_count) > MAX_WORK) {
187 spin_lock_irqsave(&async_lock, flags);
188 newcookie = next_cookie++;
189 spin_unlock_irqrestore(&async_lock, flags);
191 /* low on memory.. run synchronously */
192 ptr(data, newcookie);
197 entry->running = running;
199 spin_lock_irqsave(&async_lock, flags);
200 newcookie = entry->cookie = next_cookie++;
201 list_add_tail(&entry->list, &async_pending);
202 atomic_inc(&entry_count);
203 spin_unlock_irqrestore(&async_lock, flags);
208 async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
210 return __async_schedule(ptr, data, &async_pending);
212 EXPORT_SYMBOL_GPL(async_schedule);
214 async_cookie_t async_schedule_special(async_func_ptr *ptr, void *data, struct list_head *running)
216 return __async_schedule(ptr, data, running);
218 EXPORT_SYMBOL_GPL(async_schedule_special);
220 void async_synchronize_full(void)
223 async_synchronize_cookie(next_cookie);
224 } while (!list_empty(&async_running) || !list_empty(&async_pending));
226 EXPORT_SYMBOL_GPL(async_synchronize_full);
228 void async_synchronize_full_special(struct list_head *list)
230 async_synchronize_cookie_special(next_cookie, list);
232 EXPORT_SYMBOL_GPL(async_synchronize_full_special);
234 void async_synchronize_cookie_special(async_cookie_t cookie, struct list_head *running)
236 ktime_t starttime, delta, endtime;
238 if (initcall_debug && system_state == SYSTEM_BOOTING) {
239 printk("async_waiting @ %i\n", task_pid_nr(current));
240 starttime = ktime_get();
243 wait_event(async_done, lowest_in_progress(running) >= cookie);
245 if (initcall_debug && system_state == SYSTEM_BOOTING) {
246 endtime = ktime_get();
247 delta = ktime_sub(endtime, starttime);
249 printk("async_continuing @ %i after %lli usec\n",
250 task_pid_nr(current), ktime_to_ns(delta) >> 10);
253 EXPORT_SYMBOL_GPL(async_synchronize_cookie_special);
255 void async_synchronize_cookie(async_cookie_t cookie)
257 async_synchronize_cookie_special(cookie, &async_running);
259 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
262 static int async_thread(void *unused)
264 DECLARE_WAITQUEUE(wq, current);
265 add_wait_queue(&async_new, &wq);
267 while (!kthread_should_stop()) {
269 set_current_state(TASK_INTERRUPTIBLE);
271 * check the list head without lock.. false positives
272 * are dealt with inside run_one_entry() while holding
276 if (!list_empty(&async_pending))
279 ret = schedule_timeout(HZ);
283 * we timed out, this means we as thread are redundant.
284 * we sign off and die, but we to avoid any races there
285 * is a last-straw check to see if work snuck in.
287 atomic_dec(&thread_count);
288 wmb(); /* manager must see our departure first */
289 if (list_empty(&async_pending))
292 * woops work came in between us timing out and us
293 * signing off; we need to stay alive and keep working.
295 atomic_inc(&thread_count);
298 remove_wait_queue(&async_new, &wq);
303 static int async_manager_thread(void *unused)
305 DECLARE_WAITQUEUE(wq, current);
306 add_wait_queue(&async_new, &wq);
308 while (!kthread_should_stop()) {
311 set_current_state(TASK_INTERRUPTIBLE);
313 tc = atomic_read(&thread_count);
315 ec = atomic_read(&entry_count);
317 while (tc < ec && tc < MAX_THREADS) {
318 kthread_run(async_thread, NULL, "async/%i", tc);
319 atomic_inc(&thread_count);
325 remove_wait_queue(&async_new, &wq);
330 static int __init async_init(void)
333 kthread_run(async_manager_thread, NULL, "async/mgr");
337 static int __init setup_async(char *str)
343 __setup("fastboot", setup_async);
346 core_initcall(async_init);