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 #include "workqueue_internal.h"
61 
62 static async_cookie_t next_cookie = 1;
63 
64 #define MAX_WORK		32768
65 #define ASYNC_COOKIE_MAX	ULLONG_MAX	/* infinity cookie */
66 
67 static LIST_HEAD(async_global_pending);	/* pending from all registered doms */
68 static ASYNC_DOMAIN(async_dfl_domain);
69 static DEFINE_SPINLOCK(async_lock);
70 
71 struct async_entry {
72 	struct list_head	domain_list;
73 	struct list_head	global_list;
74 	struct work_struct	work;
75 	async_cookie_t		cookie;
76 	async_func_t		func;
77 	void			*data;
78 	struct async_domain	*domain;
79 };
80 
81 static DECLARE_WAIT_QUEUE_HEAD(async_done);
82 
83 static atomic_t entry_count;
84 
lowest_in_progress(struct async_domain * domain)85 static async_cookie_t lowest_in_progress(struct async_domain *domain)
86 {
87 	struct async_entry *first = NULL;
88 	async_cookie_t ret = ASYNC_COOKIE_MAX;
89 	unsigned long flags;
90 
91 	spin_lock_irqsave(&async_lock, flags);
92 
93 	if (domain) {
94 		if (!list_empty(&domain->pending))
95 			first = list_first_entry(&domain->pending,
96 					struct async_entry, domain_list);
97 	} else {
98 		if (!list_empty(&async_global_pending))
99 			first = list_first_entry(&async_global_pending,
100 					struct async_entry, global_list);
101 	}
102 
103 	if (first)
104 		ret = first->cookie;
105 
106 	spin_unlock_irqrestore(&async_lock, flags);
107 	return ret;
108 }
109 
110 /*
111  * pick the first pending entry and run it
112  */
async_run_entry_fn(struct work_struct * work)113 static void async_run_entry_fn(struct work_struct *work)
114 {
115 	struct async_entry *entry =
116 		container_of(work, struct async_entry, work);
117 	unsigned long flags;
118 	ktime_t uninitialized_var(calltime), delta, rettime;
119 
120 	/* 1) run (and print duration) */
121 	if (initcall_debug && system_state < SYSTEM_RUNNING) {
122 		pr_debug("calling  %lli_%pF @ %i\n",
123 			(long long)entry->cookie,
124 			entry->func, task_pid_nr(current));
125 		calltime = ktime_get();
126 	}
127 	entry->func(entry->data, entry->cookie);
128 	if (initcall_debug && system_state < SYSTEM_RUNNING) {
129 		rettime = ktime_get();
130 		delta = ktime_sub(rettime, calltime);
131 		pr_debug("initcall %lli_%pF returned 0 after %lld usecs\n",
132 			(long long)entry->cookie,
133 			entry->func,
134 			(long long)ktime_to_ns(delta) >> 10);
135 	}
136 
137 	/* 2) remove self from the pending queues */
138 	spin_lock_irqsave(&async_lock, flags);
139 	list_del_init(&entry->domain_list);
140 	list_del_init(&entry->global_list);
141 
142 	/* 3) free the entry */
143 	kfree(entry);
144 	atomic_dec(&entry_count);
145 
146 	spin_unlock_irqrestore(&async_lock, flags);
147 
148 	/* 4) wake up any waiters */
149 	wake_up(&async_done);
150 }
151 
__async_schedule(async_func_t func,void * data,struct async_domain * domain)152 static async_cookie_t __async_schedule(async_func_t func, void *data, struct async_domain *domain)
153 {
154 	struct async_entry *entry;
155 	unsigned long flags;
156 	async_cookie_t newcookie;
157 
158 	/* allow irq-off callers */
159 	entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
160 
161 	/*
162 	 * If we're out of memory or if there's too much work
163 	 * pending already, we execute synchronously.
164 	 */
165 	if (!entry || atomic_read(&entry_count) > MAX_WORK) {
166 		kfree(entry);
167 		spin_lock_irqsave(&async_lock, flags);
168 		newcookie = next_cookie++;
169 		spin_unlock_irqrestore(&async_lock, flags);
170 
171 		/* low on memory.. run synchronously */
172 		func(data, newcookie);
173 		return newcookie;
174 	}
175 	INIT_LIST_HEAD(&entry->domain_list);
176 	INIT_LIST_HEAD(&entry->global_list);
177 	INIT_WORK(&entry->work, async_run_entry_fn);
178 	entry->func = func;
179 	entry->data = data;
180 	entry->domain = domain;
181 
182 	spin_lock_irqsave(&async_lock, flags);
183 
184 	/* allocate cookie and queue */
185 	newcookie = entry->cookie = next_cookie++;
186 
187 	list_add_tail(&entry->domain_list, &domain->pending);
188 	if (domain->registered)
189 		list_add_tail(&entry->global_list, &async_global_pending);
190 
191 	atomic_inc(&entry_count);
192 	spin_unlock_irqrestore(&async_lock, flags);
193 
194 	/* mark that this task has queued an async job, used by module init */
195 	current->flags |= PF_USED_ASYNC;
196 
197 	/* schedule for execution */
198 	queue_work(system_unbound_wq, &entry->work);
199 
200 	return newcookie;
201 }
202 
203 /**
204  * async_schedule - schedule a function for asynchronous execution
205  * @func: function to execute asynchronously
206  * @data: data pointer to pass to the function
207  *
208  * Returns an async_cookie_t that may be used for checkpointing later.
209  * Note: This function may be called from atomic or non-atomic contexts.
210  */
async_schedule(async_func_t func,void * data)211 async_cookie_t async_schedule(async_func_t func, void *data)
212 {
213 	return __async_schedule(func, data, &async_dfl_domain);
214 }
215 EXPORT_SYMBOL_GPL(async_schedule);
216 
217 /**
218  * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
219  * @func: function to execute asynchronously
220  * @data: data pointer to pass to the function
221  * @domain: the domain
222  *
223  * Returns an async_cookie_t that may be used for checkpointing later.
224  * @domain may be used in the async_synchronize_*_domain() functions to
225  * wait within a certain synchronization domain rather than globally.  A
226  * synchronization domain is specified via @domain.  Note: This function
227  * may be called from atomic or non-atomic contexts.
228  */
async_schedule_domain(async_func_t func,void * data,struct async_domain * domain)229 async_cookie_t async_schedule_domain(async_func_t func, void *data,
230 				     struct async_domain *domain)
231 {
232 	return __async_schedule(func, data, domain);
233 }
234 EXPORT_SYMBOL_GPL(async_schedule_domain);
235 
236 /**
237  * async_synchronize_full - synchronize all asynchronous function calls
238  *
239  * This function waits until all asynchronous function calls have been done.
240  */
async_synchronize_full(void)241 void async_synchronize_full(void)
242 {
243 	async_synchronize_full_domain(NULL);
244 }
245 EXPORT_SYMBOL_GPL(async_synchronize_full);
246 
247 /**
248  * async_unregister_domain - ensure no more anonymous waiters on this domain
249  * @domain: idle domain to flush out of any async_synchronize_full instances
250  *
251  * async_synchronize_{cookie|full}_domain() are not flushed since callers
252  * of these routines should know the lifetime of @domain
253  *
254  * Prefer ASYNC_DOMAIN_EXCLUSIVE() declarations over flushing
255  */
async_unregister_domain(struct async_domain * domain)256 void async_unregister_domain(struct async_domain *domain)
257 {
258 	spin_lock_irq(&async_lock);
259 	WARN_ON(!domain->registered || !list_empty(&domain->pending));
260 	domain->registered = 0;
261 	spin_unlock_irq(&async_lock);
262 }
263 EXPORT_SYMBOL_GPL(async_unregister_domain);
264 
265 /**
266  * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
267  * @domain: the domain to synchronize
268  *
269  * This function waits until all asynchronous function calls for the
270  * synchronization domain specified by @domain have been done.
271  */
async_synchronize_full_domain(struct async_domain * domain)272 void async_synchronize_full_domain(struct async_domain *domain)
273 {
274 	async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain);
275 }
276 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
277 
278 /**
279  * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
280  * @cookie: async_cookie_t to use as checkpoint
281  * @domain: the domain to synchronize (%NULL for all registered domains)
282  *
283  * This function waits until all asynchronous function calls for the
284  * synchronization domain specified by @domain submitted prior to @cookie
285  * have been done.
286  */
async_synchronize_cookie_domain(async_cookie_t cookie,struct async_domain * domain)287 void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *domain)
288 {
289 	ktime_t uninitialized_var(starttime), delta, endtime;
290 
291 	if (initcall_debug && system_state < SYSTEM_RUNNING) {
292 		pr_debug("async_waiting @ %i\n", task_pid_nr(current));
293 		starttime = ktime_get();
294 	}
295 
296 	wait_event(async_done, lowest_in_progress(domain) >= cookie);
297 
298 	if (initcall_debug && system_state < SYSTEM_RUNNING) {
299 		endtime = ktime_get();
300 		delta = ktime_sub(endtime, starttime);
301 
302 		pr_debug("async_continuing @ %i after %lli usec\n",
303 			task_pid_nr(current),
304 			(long long)ktime_to_ns(delta) >> 10);
305 	}
306 }
307 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
308 
309 /**
310  * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
311  * @cookie: async_cookie_t to use as checkpoint
312  *
313  * This function waits until all asynchronous function calls prior to @cookie
314  * have been done.
315  */
async_synchronize_cookie(async_cookie_t cookie)316 void async_synchronize_cookie(async_cookie_t cookie)
317 {
318 	async_synchronize_cookie_domain(cookie, &async_dfl_domain);
319 }
320 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
321 
322 /**
323  * current_is_async - is %current an async worker task?
324  *
325  * Returns %true if %current is an async worker task.
326  */
current_is_async(void)327 bool current_is_async(void)
328 {
329 	struct worker *worker = current_wq_worker();
330 
331 	return worker && worker->current_func == async_run_entry_fn;
332 }
333 EXPORT_SYMBOL_GPL(current_is_async);
334