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/module.h>
53 #include <linux/wait.h>
54 #include <linux/sched.h>
55 #include <linux/init.h>
56 #include <linux/kthread.h>
57 #include <linux/delay.h>
58 #include <asm/atomic.h>
59
60 static async_cookie_t next_cookie = 1;
61
62 #define MAX_THREADS 256
63 #define MAX_WORK 32768
64
65 static LIST_HEAD(async_pending);
66 static LIST_HEAD(async_running);
67 static DEFINE_SPINLOCK(async_lock);
68
69 static int async_enabled = 0;
70
71 struct async_entry {
72 struct list_head list;
73 async_cookie_t cookie;
74 async_func_ptr *func;
75 void *data;
76 struct list_head *running;
77 };
78
79 static DECLARE_WAIT_QUEUE_HEAD(async_done);
80 static DECLARE_WAIT_QUEUE_HEAD(async_new);
81
82 static atomic_t entry_count;
83 static atomic_t thread_count;
84
85 extern int initcall_debug;
86
87
88 /*
89 * MUST be called with the lock held!
90 */
__lowest_in_progress(struct list_head * running)91 static async_cookie_t __lowest_in_progress(struct list_head *running)
92 {
93 struct async_entry *entry;
94 if (!list_empty(running)) {
95 entry = list_first_entry(running,
96 struct async_entry, list);
97 return entry->cookie;
98 } else if (!list_empty(&async_pending)) {
99 entry = list_first_entry(&async_pending,
100 struct async_entry, list);
101 return entry->cookie;
102 } else {
103 /* nothing in progress... next_cookie is "infinity" */
104 return next_cookie;
105 }
106
107 }
108
lowest_in_progress(struct list_head * running)109 static async_cookie_t lowest_in_progress(struct list_head *running)
110 {
111 unsigned long flags;
112 async_cookie_t ret;
113
114 spin_lock_irqsave(&async_lock, flags);
115 ret = __lowest_in_progress(running);
116 spin_unlock_irqrestore(&async_lock, flags);
117 return ret;
118 }
119 /*
120 * pick the first pending entry and run it
121 */
run_one_entry(void)122 static void run_one_entry(void)
123 {
124 unsigned long flags;
125 struct async_entry *entry;
126 ktime_t calltime, delta, rettime;
127
128 /* 1) pick one task from the pending queue */
129
130 spin_lock_irqsave(&async_lock, flags);
131 if (list_empty(&async_pending))
132 goto out;
133 entry = list_first_entry(&async_pending, struct async_entry, list);
134
135 /* 2) move it to the running queue */
136 list_move_tail(&entry->list, entry->running);
137 spin_unlock_irqrestore(&async_lock, flags);
138
139 /* 3) run it (and print duration)*/
140 if (initcall_debug && system_state == SYSTEM_BOOTING) {
141 printk("calling %lli_%pF @ %i\n", (long long)entry->cookie,
142 entry->func, task_pid_nr(current));
143 calltime = ktime_get();
144 }
145 entry->func(entry->data, entry->cookie);
146 if (initcall_debug && system_state == SYSTEM_BOOTING) {
147 rettime = ktime_get();
148 delta = ktime_sub(rettime, calltime);
149 printk("initcall %lli_%pF returned 0 after %lld usecs\n",
150 (long long)entry->cookie,
151 entry->func,
152 (long long)ktime_to_ns(delta) >> 10);
153 }
154
155 /* 4) remove it from the running queue */
156 spin_lock_irqsave(&async_lock, flags);
157 list_del(&entry->list);
158
159 /* 5) free the entry */
160 kfree(entry);
161 atomic_dec(&entry_count);
162
163 spin_unlock_irqrestore(&async_lock, flags);
164
165 /* 6) wake up any waiters. */
166 wake_up(&async_done);
167 return;
168
169 out:
170 spin_unlock_irqrestore(&async_lock, flags);
171 }
172
173
__async_schedule(async_func_ptr * ptr,void * data,struct list_head * running)174 static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running)
175 {
176 struct async_entry *entry;
177 unsigned long flags;
178 async_cookie_t newcookie;
179
180
181 /* allow irq-off callers */
182 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
183
184 /*
185 * If we're out of memory or if there's too much work
186 * pending already, we execute synchronously.
187 */
188 if (!async_enabled || !entry || atomic_read(&entry_count) > MAX_WORK) {
189 kfree(entry);
190 spin_lock_irqsave(&async_lock, flags);
191 newcookie = next_cookie++;
192 spin_unlock_irqrestore(&async_lock, flags);
193
194 /* low on memory.. run synchronously */
195 ptr(data, newcookie);
196 return newcookie;
197 }
198 entry->func = ptr;
199 entry->data = data;
200 entry->running = running;
201
202 spin_lock_irqsave(&async_lock, flags);
203 newcookie = entry->cookie = next_cookie++;
204 list_add_tail(&entry->list, &async_pending);
205 atomic_inc(&entry_count);
206 spin_unlock_irqrestore(&async_lock, flags);
207 wake_up(&async_new);
208 return newcookie;
209 }
210
211 /**
212 * async_schedule - schedule a function for asynchronous execution
213 * @ptr: function to execute asynchronously
214 * @data: data pointer to pass to the function
215 *
216 * Returns an async_cookie_t that may be used for checkpointing later.
217 * Note: This function may be called from atomic or non-atomic contexts.
218 */
async_schedule(async_func_ptr * ptr,void * data)219 async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
220 {
221 return __async_schedule(ptr, data, &async_running);
222 }
223 EXPORT_SYMBOL_GPL(async_schedule);
224
225 /**
226 * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
227 * @ptr: function to execute asynchronously
228 * @data: data pointer to pass to the function
229 * @running: running list for the domain
230 *
231 * Returns an async_cookie_t that may be used for checkpointing later.
232 * @running may be used in the async_synchronize_*_domain() functions
233 * to wait within a certain synchronization domain rather than globally.
234 * A synchronization domain is specified via the running queue @running to use.
235 * Note: This function may be called from atomic or non-atomic contexts.
236 */
async_schedule_domain(async_func_ptr * ptr,void * data,struct list_head * running)237 async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data,
238 struct list_head *running)
239 {
240 return __async_schedule(ptr, data, running);
241 }
242 EXPORT_SYMBOL_GPL(async_schedule_domain);
243
244 /**
245 * async_synchronize_full - synchronize all asynchronous function calls
246 *
247 * This function waits until all asynchronous function calls have been done.
248 */
async_synchronize_full(void)249 void async_synchronize_full(void)
250 {
251 do {
252 async_synchronize_cookie(next_cookie);
253 } while (!list_empty(&async_running) || !list_empty(&async_pending));
254 }
255 EXPORT_SYMBOL_GPL(async_synchronize_full);
256
257 /**
258 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
259 * @list: running list to synchronize on
260 *
261 * This function waits until all asynchronous function calls for the
262 * synchronization domain specified by the running list @list have been done.
263 */
async_synchronize_full_domain(struct list_head * list)264 void async_synchronize_full_domain(struct list_head *list)
265 {
266 async_synchronize_cookie_domain(next_cookie, list);
267 }
268 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
269
270 /**
271 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
272 * @cookie: async_cookie_t to use as checkpoint
273 * @running: running list to synchronize on
274 *
275 * This function waits until all asynchronous function calls for the
276 * synchronization domain specified by the running list @list submitted
277 * prior to @cookie have been done.
278 */
async_synchronize_cookie_domain(async_cookie_t cookie,struct list_head * running)279 void async_synchronize_cookie_domain(async_cookie_t cookie,
280 struct list_head *running)
281 {
282 ktime_t starttime, delta, endtime;
283
284 if (initcall_debug && system_state == SYSTEM_BOOTING) {
285 printk("async_waiting @ %i\n", task_pid_nr(current));
286 starttime = ktime_get();
287 }
288
289 wait_event(async_done, lowest_in_progress(running) >= cookie);
290
291 if (initcall_debug && system_state == SYSTEM_BOOTING) {
292 endtime = ktime_get();
293 delta = ktime_sub(endtime, starttime);
294
295 printk("async_continuing @ %i after %lli usec\n",
296 task_pid_nr(current),
297 (long long)ktime_to_ns(delta) >> 10);
298 }
299 }
300 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
301
302 /**
303 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
304 * @cookie: async_cookie_t to use as checkpoint
305 *
306 * This function waits until all asynchronous function calls prior to @cookie
307 * have been done.
308 */
async_synchronize_cookie(async_cookie_t cookie)309 void async_synchronize_cookie(async_cookie_t cookie)
310 {
311 async_synchronize_cookie_domain(cookie, &async_running);
312 }
313 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
314
315
async_thread(void * unused)316 static int async_thread(void *unused)
317 {
318 DECLARE_WAITQUEUE(wq, current);
319 add_wait_queue(&async_new, &wq);
320
321 while (!kthread_should_stop()) {
322 int ret = HZ;
323 set_current_state(TASK_INTERRUPTIBLE);
324 /*
325 * check the list head without lock.. false positives
326 * are dealt with inside run_one_entry() while holding
327 * the lock.
328 */
329 rmb();
330 if (!list_empty(&async_pending))
331 run_one_entry();
332 else
333 ret = schedule_timeout(HZ);
334
335 if (ret == 0) {
336 /*
337 * we timed out, this means we as thread are redundant.
338 * we sign off and die, but we to avoid any races there
339 * is a last-straw check to see if work snuck in.
340 */
341 atomic_dec(&thread_count);
342 wmb(); /* manager must see our departure first */
343 if (list_empty(&async_pending))
344 break;
345 /*
346 * woops work came in between us timing out and us
347 * signing off; we need to stay alive and keep working.
348 */
349 atomic_inc(&thread_count);
350 }
351 }
352 remove_wait_queue(&async_new, &wq);
353
354 return 0;
355 }
356
async_manager_thread(void * unused)357 static int async_manager_thread(void *unused)
358 {
359 DECLARE_WAITQUEUE(wq, current);
360 add_wait_queue(&async_new, &wq);
361
362 while (!kthread_should_stop()) {
363 int tc, ec;
364
365 set_current_state(TASK_INTERRUPTIBLE);
366
367 tc = atomic_read(&thread_count);
368 rmb();
369 ec = atomic_read(&entry_count);
370
371 while (tc < ec && tc < MAX_THREADS) {
372 if (IS_ERR(kthread_run(async_thread, NULL, "async/%i",
373 tc))) {
374 msleep(100);
375 continue;
376 }
377 atomic_inc(&thread_count);
378 tc++;
379 }
380
381 schedule();
382 }
383 remove_wait_queue(&async_new, &wq);
384
385 return 0;
386 }
387
async_init(void)388 static int __init async_init(void)
389 {
390 if (async_enabled)
391 if (IS_ERR(kthread_run(async_manager_thread, NULL,
392 "async/mgr")))
393 async_enabled = 0;
394 return 0;
395 }
396
setup_async(char * str)397 static int __init setup_async(char *str)
398 {
399 async_enabled = 1;
400 return 1;
401 }
402
403 __setup("fastboot", setup_async);
404
405
406 core_initcall(async_init);
407