1 // SPDX-License-Identifier: GPL-2.0-only
3 * async.c: Asynchronous function calls for boot performance
5 * (C) Copyright 2009 Intel Corporation
6 * Author: Arjan van de Ven <arjan@linux.intel.com>
12 Goals and Theory of Operation
14 The primary goal of this feature is to reduce the kernel boot time,
15 by doing various independent hardware delays and discovery operations
16 decoupled and not strictly serialized.
18 More specifically, the asynchronous function call concept allows
19 certain operations (primarily during system boot) to happen
20 asynchronously, out of order, while these operations still
21 have their externally visible parts happen sequentially and in-order.
22 (not unlike how out-of-order CPUs retire their instructions in order)
24 Key to the asynchronous function call implementation is the concept of
25 a "sequence cookie" (which, although it has an abstracted type, can be
26 thought of as a monotonically incrementing number).
28 The async core will assign each scheduled event such a sequence cookie and
29 pass this to the called functions.
31 The asynchronously called function should before doing a globally visible
32 operation, such as registering device numbers, call the
33 async_synchronize_cookie() function and pass in its own cookie. The
34 async_synchronize_cookie() function will make sure that all asynchronous
35 operations that were scheduled prior to the operation corresponding with the
36 cookie have completed.
38 Subsystem/driver initialization code that scheduled asynchronous probe
39 functions, but which shares global resources with other drivers/subsystems
40 that do not use the asynchronous call feature, need to do a full
41 synchronization with the async_synchronize_full() function, before returning
42 from their init function. This is to maintain strict ordering between the
43 asynchronous and synchronous parts of the kernel.
47 #include <linux/async.h>
48 #include <linux/atomic.h>
49 #include <linux/ktime.h>
50 #include <linux/export.h>
51 #include <linux/wait.h>
52 #include <linux/sched.h>
53 #include <linux/slab.h>
54 #include <linux/workqueue.h>
56 #include "workqueue_internal.h"
58 static async_cookie_t next_cookie = 1;
60 #define MAX_WORK 32768
61 #define ASYNC_COOKIE_MAX ULLONG_MAX /* infinity cookie */
63 static LIST_HEAD(async_global_pending); /* pending from all registered doms */
64 static ASYNC_DOMAIN(async_dfl_domain);
65 static DEFINE_SPINLOCK(async_lock);
68 struct list_head domain_list;
69 struct list_head global_list;
70 struct work_struct work;
71 async_cookie_t cookie;
74 struct async_domain *domain;
77 static DECLARE_WAIT_QUEUE_HEAD(async_done);
79 static atomic_t entry_count;
81 static long long microseconds_since(ktime_t start)
83 ktime_t now = ktime_get();
84 return ktime_to_ns(ktime_sub(now, start)) >> 10;
87 static async_cookie_t lowest_in_progress(struct async_domain *domain)
89 struct async_entry *first = NULL;
90 async_cookie_t ret = ASYNC_COOKIE_MAX;
93 spin_lock_irqsave(&async_lock, flags);
96 if (!list_empty(&domain->pending))
97 first = list_first_entry(&domain->pending,
98 struct async_entry, domain_list);
100 if (!list_empty(&async_global_pending))
101 first = list_first_entry(&async_global_pending,
102 struct async_entry, global_list);
108 spin_unlock_irqrestore(&async_lock, flags);
113 * pick the first pending entry and run it
115 static void async_run_entry_fn(struct work_struct *work)
117 struct async_entry *entry =
118 container_of(work, struct async_entry, work);
122 /* 1) run (and print duration) */
123 pr_debug("calling %lli_%pS @ %i\n", (long long)entry->cookie,
124 entry->func, task_pid_nr(current));
125 calltime = ktime_get();
127 entry->func(entry->data, entry->cookie);
129 pr_debug("initcall %lli_%pS returned after %lld usecs\n",
130 (long long)entry->cookie, entry->func,
131 microseconds_since(calltime));
133 /* 2) remove self from the pending queues */
134 spin_lock_irqsave(&async_lock, flags);
135 list_del_init(&entry->domain_list);
136 list_del_init(&entry->global_list);
138 /* 3) free the entry */
140 atomic_dec(&entry_count);
142 spin_unlock_irqrestore(&async_lock, flags);
144 /* 4) wake up any waiters */
145 wake_up(&async_done);
148 static async_cookie_t __async_schedule_node_domain(async_func_t func,
149 void *data, int node,
150 struct async_domain *domain,
151 struct async_entry *entry)
153 async_cookie_t newcookie;
156 INIT_LIST_HEAD(&entry->domain_list);
157 INIT_LIST_HEAD(&entry->global_list);
158 INIT_WORK(&entry->work, async_run_entry_fn);
161 entry->domain = domain;
163 spin_lock_irqsave(&async_lock, flags);
165 /* allocate cookie and queue */
166 newcookie = entry->cookie = next_cookie++;
168 list_add_tail(&entry->domain_list, &domain->pending);
169 if (domain->registered)
170 list_add_tail(&entry->global_list, &async_global_pending);
172 atomic_inc(&entry_count);
173 spin_unlock_irqrestore(&async_lock, flags);
175 /* schedule for execution */
176 queue_work_node(node, system_unbound_wq, &entry->work);
182 * async_schedule_node_domain - NUMA specific version of async_schedule_domain
183 * @func: function to execute asynchronously
184 * @data: data pointer to pass to the function
185 * @node: NUMA node that we want to schedule this on or close to
186 * @domain: the domain
188 * Returns an async_cookie_t that may be used for checkpointing later.
189 * @domain may be used in the async_synchronize_*_domain() functions to
190 * wait within a certain synchronization domain rather than globally.
192 * Note: This function may be called from atomic or non-atomic contexts.
194 * The node requested will be honored on a best effort basis. If the node
195 * has no CPUs associated with it then the work is distributed among all
198 async_cookie_t async_schedule_node_domain(async_func_t func, void *data,
199 int node, struct async_domain *domain)
201 struct async_entry *entry;
203 async_cookie_t newcookie;
205 /* allow irq-off callers */
206 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
209 * If we're out of memory or if there's too much work
210 * pending already, we execute synchronously.
212 if (!entry || atomic_read(&entry_count) > MAX_WORK) {
214 spin_lock_irqsave(&async_lock, flags);
215 newcookie = next_cookie++;
216 spin_unlock_irqrestore(&async_lock, flags);
218 /* low on memory.. run synchronously */
219 func(data, newcookie);
223 return __async_schedule_node_domain(func, data, node, domain, entry);
225 EXPORT_SYMBOL_GPL(async_schedule_node_domain);
228 * async_schedule_node - NUMA specific version of async_schedule
229 * @func: function to execute asynchronously
230 * @data: data pointer to pass to the function
231 * @node: NUMA node that we want to schedule this on or close to
233 * Returns an async_cookie_t that may be used for checkpointing later.
234 * Note: This function may be called from atomic or non-atomic contexts.
236 * The node requested will be honored on a best effort basis. If the node
237 * has no CPUs associated with it then the work is distributed among all
240 async_cookie_t async_schedule_node(async_func_t func, void *data, int node)
242 return async_schedule_node_domain(func, data, node, &async_dfl_domain);
244 EXPORT_SYMBOL_GPL(async_schedule_node);
247 * async_schedule_dev_nocall - A simplified variant of async_schedule_dev()
248 * @func: function to execute asynchronously
249 * @dev: device argument to be passed to function
251 * @dev is used as both the argument for the function and to provide NUMA
252 * context for where to run the function.
254 * If the asynchronous execution of @func is scheduled successfully, return
255 * true. Otherwise, do nothing and return false, unlike async_schedule_dev()
256 * that will run the function synchronously then.
258 bool async_schedule_dev_nocall(async_func_t func, struct device *dev)
260 struct async_entry *entry;
262 entry = kzalloc(sizeof(struct async_entry), GFP_KERNEL);
264 /* Give up if there is no memory or too much work. */
265 if (!entry || atomic_read(&entry_count) > MAX_WORK) {
270 __async_schedule_node_domain(func, dev, dev_to_node(dev),
271 &async_dfl_domain, entry);
276 * async_synchronize_full - synchronize all asynchronous function calls
278 * This function waits until all asynchronous function calls have been done.
280 void async_synchronize_full(void)
282 async_synchronize_full_domain(NULL);
284 EXPORT_SYMBOL_GPL(async_synchronize_full);
287 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
288 * @domain: the domain to synchronize
290 * This function waits until all asynchronous function calls for the
291 * synchronization domain specified by @domain have been done.
293 void async_synchronize_full_domain(struct async_domain *domain)
295 async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain);
297 EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
300 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
301 * @cookie: async_cookie_t to use as checkpoint
302 * @domain: the domain to synchronize (%NULL for all registered domains)
304 * This function waits until all asynchronous function calls for the
305 * synchronization domain specified by @domain submitted prior to @cookie
308 void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *domain)
312 pr_debug("async_waiting @ %i\n", task_pid_nr(current));
313 starttime = ktime_get();
315 wait_event(async_done, lowest_in_progress(domain) >= cookie);
317 pr_debug("async_continuing @ %i after %lli usec\n", task_pid_nr(current),
318 microseconds_since(starttime));
320 EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
323 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
324 * @cookie: async_cookie_t to use as checkpoint
326 * This function waits until all asynchronous function calls prior to @cookie
329 void async_synchronize_cookie(async_cookie_t cookie)
331 async_synchronize_cookie_domain(cookie, &async_dfl_domain);
333 EXPORT_SYMBOL_GPL(async_synchronize_cookie);
336 * current_is_async - is %current an async worker task?
338 * Returns %true if %current is an async worker task.
340 bool current_is_async(void)
342 struct worker *worker = current_wq_worker();
344 return worker && worker->current_func == async_run_entry_fn;
346 EXPORT_SYMBOL_GPL(current_is_async);