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/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>
60 #include "workqueue_internal.h"
62 static async_cookie_t next_cookie = 1;
64 #define MAX_WORK 32768
65 #define ASYNC_COOKIE_MAX ULLONG_MAX /* infinity cookie */
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);
72 struct list_head domain_list;
73 struct list_head global_list;
74 struct work_struct work;
75 async_cookie_t cookie;
78 struct async_domain *domain;
81 static DECLARE_WAIT_QUEUE_HEAD(async_done);
83 static atomic_t entry_count;
85 static async_cookie_t lowest_in_progress(struct async_domain *domain)
87 struct async_entry *first = NULL;
88 async_cookie_t ret = ASYNC_COOKIE_MAX;
91 spin_lock_irqsave(&async_lock, flags);
94 if (!list_empty(&domain->pending))
95 first = list_first_entry(&domain->pending,
96 struct async_entry, domain_list);
98 if (!list_empty(&async_global_pending))
99 first = list_first_entry(&async_global_pending,
100 struct async_entry, global_list);
106 spin_unlock_irqrestore(&async_lock, flags);
111 * pick the first pending entry and run it
113 static void async_run_entry_fn(struct work_struct *work)
115 struct async_entry *entry =
116 container_of(work, struct async_entry, work);
118 ktime_t uninitialized_var(calltime), delta, rettime;
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();
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,
134 (long long)ktime_to_ns(delta) >> 10);
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);
142 /* 3) free the entry */
144 atomic_dec(&entry_count);
146 spin_unlock_irqrestore(&async_lock, flags);
148 /* 4) wake up any waiters */
149 wake_up(&async_done);
153 * async_schedule_node_domain - NUMA specific version of async_schedule_domain
154 * @func: function to execute asynchronously
155 * @data: data pointer to pass to the function
156 * @node: NUMA node that we want to schedule this on or close to
157 * @domain: the domain
159 * Returns an async_cookie_t that may be used for checkpointing later.
160 * @domain may be used in the async_synchronize_*_domain() functions to
161 * wait within a certain synchronization domain rather than globally.
163 * Note: This function may be called from atomic or non-atomic contexts.
165 * The node requested will be honored on a best effort basis. If the node
166 * has no CPUs associated with it then the work is distributed among all
169 async_cookie_t async_schedule_node_domain(async_func_t func, void *data,
170 int node, struct async_domain *domain)
172 struct async_entry *entry;
174 async_cookie_t newcookie;
176 /* allow irq-off callers */
177 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
180 * If we're out of memory or if there's too much work
181 * pending already, we execute synchronously.
183 if (!entry || atomic_read(&entry_count) > MAX_WORK) {
185 spin_lock_irqsave(&async_lock, flags);
186 newcookie = next_cookie++;
187 spin_unlock_irqrestore(&async_lock, flags);
189 /* low on memory.. run synchronously */
190 func(data, newcookie);
193 INIT_LIST_HEAD(&entry->domain_list);
194 INIT_LIST_HEAD(&entry->global_list);
195 INIT_WORK(&entry->work, async_run_entry_fn);
198 entry->domain = domain;
200 spin_lock_irqsave(&async_lock, flags);
202 /* allocate cookie and queue */
203 newcookie = entry->cookie = next_cookie++;
205 list_add_tail(&entry->domain_list, &domain->pending);
206 if (domain->registered)
207 list_add_tail(&entry->global_list, &async_global_pending);
209 atomic_inc(&entry_count);
210 spin_unlock_irqrestore(&async_lock, flags);
212 /* mark that this task has queued an async job, used by module init */
213 current->flags |= PF_USED_ASYNC;
215 /* schedule for execution */
216 queue_work_node(node, system_unbound_wq, &entry->work);
220 EXPORT_SYMBOL_GPL(async_schedule_node_domain);
223 * async_schedule_node - NUMA specific version of async_schedule
224 * @func: function to execute asynchronously
225 * @data: data pointer to pass to the function
226 * @node: NUMA node that we want to schedule this on or close to
228 * Returns an async_cookie_t that may be used for checkpointing later.
229 * Note: This function may be called from atomic or non-atomic contexts.
231 * The node requested will be honored on a best effort basis. If the node
232 * has no CPUs associated with it then the work is distributed among all
235 async_cookie_t async_schedule_node(async_func_t func, void *data, int node)
237 return async_schedule_node_domain(func, data, node, &async_dfl_domain);
239 EXPORT_SYMBOL_GPL(async_schedule_node);
242 * async_synchronize_full - synchronize all asynchronous function calls
244 * This function waits until all asynchronous function calls have been done.
246 void async_synchronize_full(void)
248 async_synchronize_full_domain(NULL);
250 EXPORT_SYMBOL_GPL(async_synchronize_full);
253 * async_unregister_domain - ensure no more anonymous waiters on this domain
254 * @domain: idle domain to flush out of any async_synchronize_full instances
256 * async_synchronize_{cookie|full}_domain() are not flushed since callers
257 * of these routines should know the lifetime of @domain
259 * Prefer ASYNC_DOMAIN_EXCLUSIVE() declarations over flushing
261 void async_unregister_domain(struct async_domain *domain)
263 spin_lock_irq(&async_lock);
264 WARN_ON(!domain->registered || !list_empty(&domain->pending));
265 domain->registered = 0;
266 spin_unlock_irq(&async_lock);
268 EXPORT_SYMBOL_GPL(async_unregister_domain);
271 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
272 * @domain: the domain to synchronize
274 * This function waits until all asynchronous function calls for the
275 * synchronization domain specified by @domain have been done.
277 void async_synchronize_full_domain(struct async_domain *domain)
279 async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain);
281 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
284 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
285 * @cookie: async_cookie_t to use as checkpoint
286 * @domain: the domain to synchronize (%NULL for all registered domains)
288 * This function waits until all asynchronous function calls for the
289 * synchronization domain specified by @domain submitted prior to @cookie
292 void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *domain)
294 ktime_t uninitialized_var(starttime), delta, endtime;
296 if (initcall_debug && system_state < SYSTEM_RUNNING) {
297 pr_debug("async_waiting @ %i\n", task_pid_nr(current));
298 starttime = ktime_get();
301 wait_event(async_done, lowest_in_progress(domain) >= cookie);
303 if (initcall_debug && system_state < SYSTEM_RUNNING) {
304 endtime = ktime_get();
305 delta = ktime_sub(endtime, starttime);
307 pr_debug("async_continuing @ %i after %lli usec\n",
308 task_pid_nr(current),
309 (long long)ktime_to_ns(delta) >> 10);
312 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
315 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
316 * @cookie: async_cookie_t to use as checkpoint
318 * This function waits until all asynchronous function calls prior to @cookie
321 void async_synchronize_cookie(async_cookie_t cookie)
323 async_synchronize_cookie_domain(cookie, &async_dfl_domain);
325 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
328 * current_is_async - is %current an async worker task?
330 * Returns %true if %current is an async worker task.
332 bool current_is_async(void)
334 struct worker *worker = current_wq_worker();
336 return worker && worker->current_func == async_run_entry_fn;
338 EXPORT_SYMBOL_GPL(current_is_async);