Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
v3.1
 
  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
 16Goals and Theory of Operation
 17
 18The primary goal of this feature is to reduce the kernel boot time,
 19by doing various independent hardware delays and discovery operations
 20decoupled and not strictly serialized.
 21
 22More specifically, the asynchronous function call concept allows
 23certain operations (primarily during system boot) to happen
 24asynchronously, out of order, while these operations still
 25have their externally visible parts happen sequentially and in-order.
 26(not unlike how out-of-order CPUs retire their instructions in order)
 27
 28Key to the asynchronous function call implementation is the concept of
 29a "sequence cookie" (which, although it has an abstracted type, can be
 30thought of as a monotonically incrementing number).
 31
 32The async core will assign each scheduled event such a sequence cookie and
 33pass this to the called functions.
 34
 35The asynchronously called function should before doing a globally visible
 36operation, such as registering device numbers, call the
 37async_synchronize_cookie() function and pass in its own cookie. The
 38async_synchronize_cookie() function will make sure that all asynchronous
 39operations that were scheduled prior to the operation corresponding with the
 40cookie have completed.
 41
 42Subsystem/driver initialization code that scheduled asynchronous probe
 43functions, but which shares global resources with other drivers/subsystems
 44that do not use the asynchronous call feature, need to do a full
 45synchronization with the async_synchronize_full() function, before returning
 46from their init function. This is to maintain strict ordering between the
 47asynchronous 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/module.h>
 55#include <linux/wait.h>
 56#include <linux/sched.h>
 57#include <linux/slab.h>
 
 58#include <linux/workqueue.h>
 59
 
 
 60static async_cookie_t next_cookie = 1;
 61
 62#define MAX_WORK	32768
 
 63
 64static LIST_HEAD(async_pending);
 65static LIST_HEAD(async_running);
 66static DEFINE_SPINLOCK(async_lock);
 
 67
 68struct async_entry {
 69	struct list_head	list;
 
 70	struct work_struct	work;
 71	async_cookie_t		cookie;
 72	async_func_ptr		*func;
 73	void			*data;
 74	struct list_head	*running;
 75};
 76
 77static DECLARE_WAIT_QUEUE_HEAD(async_done);
 78
 79static atomic_t entry_count;
 80
 81extern int initcall_debug;
 82
 83
 84/*
 85 * MUST be called with the lock held!
 86 */
 87static async_cookie_t  __lowest_in_progress(struct list_head *running)
 88{
 89	struct async_entry *entry;
 90
 91	if (!list_empty(running)) {
 92		entry = list_first_entry(running,
 93			struct async_entry, list);
 94		return entry->cookie;
 95	}
 96
 97	list_for_each_entry(entry, &async_pending, list)
 98		if (entry->running == running)
 99			return entry->cookie;
100
101	return next_cookie;	/* "infinity" value */
102}
103
104static async_cookie_t  lowest_in_progress(struct list_head *running)
105{
 
 
106	unsigned long flags;
107	async_cookie_t ret;
108
109	spin_lock_irqsave(&async_lock, flags);
110	ret = __lowest_in_progress(running);
 
 
 
 
 
 
 
 
 
 
 
 
 
111	spin_unlock_irqrestore(&async_lock, flags);
112	return ret;
113}
114
115/*
116 * pick the first pending entry and run it
117 */
118static void async_run_entry_fn(struct work_struct *work)
119{
120	struct async_entry *entry =
121		container_of(work, struct async_entry, work);
122	unsigned long flags;
123	ktime_t calltime, delta, rettime;
124
125	/* 1) move self to the running queue */
126	spin_lock_irqsave(&async_lock, flags);
127	list_move_tail(&entry->list, entry->running);
128	spin_unlock_irqrestore(&async_lock, flags);
129
130	/* 2) run (and print duration) */
131	if (initcall_debug && system_state == SYSTEM_BOOTING) {
132		printk(KERN_DEBUG "calling  %lli_%pF @ %i\n",
133			(long long)entry->cookie,
134			entry->func, task_pid_nr(current));
135		calltime = ktime_get();
136	}
137	entry->func(entry->data, entry->cookie);
138	if (initcall_debug && system_state == SYSTEM_BOOTING) {
139		rettime = ktime_get();
140		delta = ktime_sub(rettime, calltime);
141		printk(KERN_DEBUG "initcall %lli_%pF returned 0 after %lld usecs\n",
142			(long long)entry->cookie,
143			entry->func,
144			(long long)ktime_to_ns(delta) >> 10);
145	}
146
147	/* 3) remove self from the running queue */
 
 
 
 
148	spin_lock_irqsave(&async_lock, flags);
149	list_del(&entry->list);
 
150
151	/* 4) free the entry */
152	kfree(entry);
153	atomic_dec(&entry_count);
154
155	spin_unlock_irqrestore(&async_lock, flags);
156
157	/* 5) wake up any waiters */
158	wake_up(&async_done);
159}
160
161static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
162{
163	struct async_entry *entry;
164	unsigned long flags;
165	async_cookie_t newcookie;
166
167	/* allow irq-off callers */
168	entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
169
170	/*
171	 * If we're out of memory or if there's too much work
172	 * pending already, we execute synchronously.
173	 */
174	if (!entry || atomic_read(&entry_count) > MAX_WORK) {
175		kfree(entry);
176		spin_lock_irqsave(&async_lock, flags);
177		newcookie = next_cookie++;
178		spin_unlock_irqrestore(&async_lock, flags);
179
180		/* low on memory.. run synchronously */
181		ptr(data, newcookie);
182		return newcookie;
183	}
184	INIT_WORK(&entry->work, async_run_entry_fn);
185	entry->func = ptr;
186	entry->data = data;
187	entry->running = running;
188
189	spin_lock_irqsave(&async_lock, flags);
190	newcookie = entry->cookie = next_cookie++;
191	list_add_tail(&entry->list, &async_pending);
192	atomic_inc(&entry_count);
193	spin_unlock_irqrestore(&async_lock, flags);
194
195	/* schedule for execution */
196	queue_work(system_unbound_wq, &entry->work);
197
198	return newcookie;
199}
 
200
201/**
202 * async_schedule - schedule a function for asynchronous execution
203 * @ptr: function to execute asynchronously
204 * @data: data pointer to pass to the function
 
205 *
206 * Returns an async_cookie_t that may be used for checkpointing later.
207 * Note: This function may be called from atomic or non-atomic contexts.
 
 
 
 
208 */
209async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
210{
211	return __async_schedule(ptr, data, &async_running);
212}
213EXPORT_SYMBOL_GPL(async_schedule);
214
215/**
216 * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
217 * @ptr: function to execute asynchronously
218 * @data: data pointer to pass to the function
219 * @running: running list for the domain
220 *
221 * Returns an async_cookie_t that may be used for checkpointing later.
222 * @running may be used in the async_synchronize_*_domain() functions
223 * to wait within a certain synchronization domain rather than globally.
224 * A synchronization domain is specified via the running queue @running to use.
225 * Note: This function may be called from atomic or non-atomic contexts.
 
226 */
227async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data,
228				     struct list_head *running)
229{
230	return __async_schedule(ptr, data, running);
 
 
 
 
 
 
 
 
 
 
 
 
231}
232EXPORT_SYMBOL_GPL(async_schedule_domain);
233
234/**
235 * async_synchronize_full - synchronize all asynchronous function calls
236 *
237 * This function waits until all asynchronous function calls have been done.
238 */
239void async_synchronize_full(void)
240{
241	do {
242		async_synchronize_cookie(next_cookie);
243	} while (!list_empty(&async_running) || !list_empty(&async_pending));
244}
245EXPORT_SYMBOL_GPL(async_synchronize_full);
246
247/**
248 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
249 * @list: running list to synchronize on
250 *
251 * This function waits until all asynchronous function calls for the
252 * synchronization domain specified by the running list @list have been done.
253 */
254void async_synchronize_full_domain(struct list_head *list)
255{
256	async_synchronize_cookie_domain(next_cookie, list);
257}
258EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
259
260/**
261 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
262 * @cookie: async_cookie_t to use as checkpoint
263 * @running: running list to synchronize on
264 *
265 * This function waits until all asynchronous function calls for the
266 * synchronization domain specified by the running list @list submitted
267 * prior to @cookie have been done.
268 */
269void async_synchronize_cookie_domain(async_cookie_t cookie,
270				     struct list_head *running)
271{
272	ktime_t starttime, delta, endtime;
273
274	if (initcall_debug && system_state == SYSTEM_BOOTING) {
275		printk(KERN_DEBUG "async_waiting @ %i\n", task_pid_nr(current));
276		starttime = ktime_get();
277	}
278
279	wait_event(async_done, lowest_in_progress(running) >= cookie);
280
281	if (initcall_debug && system_state == SYSTEM_BOOTING) {
282		endtime = ktime_get();
283		delta = ktime_sub(endtime, starttime);
284
285		printk(KERN_DEBUG "async_continuing @ %i after %lli usec\n",
286			task_pid_nr(current),
287			(long long)ktime_to_ns(delta) >> 10);
288	}
289}
290EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
291
292/**
293 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
294 * @cookie: async_cookie_t to use as checkpoint
295 *
296 * This function waits until all asynchronous function calls prior to @cookie
297 * have been done.
298 */
299void async_synchronize_cookie(async_cookie_t cookie)
300{
301	async_synchronize_cookie_domain(cookie, &async_running);
302}
303EXPORT_SYMBOL_GPL(async_synchronize_cookie);
v6.9.4
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * async.c: Asynchronous function calls for boot performance
  4 *
  5 * (C) Copyright 2009 Intel Corporation
  6 * Author: Arjan van de Ven <arjan@linux.intel.com>
 
 
 
 
 
  7 */
  8
  9
 10/*
 11
 12Goals and Theory of Operation
 13
 14The primary goal of this feature is to reduce the kernel boot time,
 15by doing various independent hardware delays and discovery operations
 16decoupled and not strictly serialized.
 17
 18More specifically, the asynchronous function call concept allows
 19certain operations (primarily during system boot) to happen
 20asynchronously, out of order, while these operations still
 21have their externally visible parts happen sequentially and in-order.
 22(not unlike how out-of-order CPUs retire their instructions in order)
 23
 24Key to the asynchronous function call implementation is the concept of
 25a "sequence cookie" (which, although it has an abstracted type, can be
 26thought of as a monotonically incrementing number).
 27
 28The async core will assign each scheduled event such a sequence cookie and
 29pass this to the called functions.
 30
 31The asynchronously called function should before doing a globally visible
 32operation, such as registering device numbers, call the
 33async_synchronize_cookie() function and pass in its own cookie. The
 34async_synchronize_cookie() function will make sure that all asynchronous
 35operations that were scheduled prior to the operation corresponding with the
 36cookie have completed.
 37
 38Subsystem/driver initialization code that scheduled asynchronous probe
 39functions, but which shares global resources with other drivers/subsystems
 40that do not use the asynchronous call feature, need to do a full
 41synchronization with the async_synchronize_full() function, before returning
 42from their init function. This is to maintain strict ordering between the
 43asynchronous and synchronous parts of the kernel.
 44
 45*/
 46
 47#include <linux/async.h>
 48#include <linux/atomic.h>
 49#include <linux/export.h>
 50#include <linux/ktime.h>
 51#include <linux/pid.h>
 
 52#include <linux/sched.h>
 53#include <linux/slab.h>
 54#include <linux/wait.h>
 55#include <linux/workqueue.h>
 56
 57#include "workqueue_internal.h"
 58
 59static async_cookie_t next_cookie = 1;
 60
 61#define MAX_WORK		32768
 62#define ASYNC_COOKIE_MAX	ULLONG_MAX	/* infinity cookie */
 63
 64static LIST_HEAD(async_global_pending);	/* pending from all registered doms */
 65static ASYNC_DOMAIN(async_dfl_domain);
 66static DEFINE_SPINLOCK(async_lock);
 67static struct workqueue_struct *async_wq;
 68
 69struct async_entry {
 70	struct list_head	domain_list;
 71	struct list_head	global_list;
 72	struct work_struct	work;
 73	async_cookie_t		cookie;
 74	async_func_t		func;
 75	void			*data;
 76	struct async_domain	*domain;
 77};
 78
 79static DECLARE_WAIT_QUEUE_HEAD(async_done);
 80
 81static atomic_t entry_count;
 82
 83static long long microseconds_since(ktime_t start)
 
 
 
 
 
 
 84{
 85	ktime_t now = ktime_get();
 86	return ktime_to_ns(ktime_sub(now, start)) >> 10;
 
 
 
 
 
 
 
 
 
 
 
 87}
 88
 89static async_cookie_t lowest_in_progress(struct async_domain *domain)
 90{
 91	struct async_entry *first = NULL;
 92	async_cookie_t ret = ASYNC_COOKIE_MAX;
 93	unsigned long flags;
 
 94
 95	spin_lock_irqsave(&async_lock, flags);
 96
 97	if (domain) {
 98		if (!list_empty(&domain->pending))
 99			first = list_first_entry(&domain->pending,
100					struct async_entry, domain_list);
101	} else {
102		if (!list_empty(&async_global_pending))
103			first = list_first_entry(&async_global_pending,
104					struct async_entry, global_list);
105	}
106
107	if (first)
108		ret = first->cookie;
109
110	spin_unlock_irqrestore(&async_lock, flags);
111	return ret;
112}
113
114/*
115 * pick the first pending entry and run it
116 */
117static void async_run_entry_fn(struct work_struct *work)
118{
119	struct async_entry *entry =
120		container_of(work, struct async_entry, work);
121	unsigned long flags;
122	ktime_t calltime;
123
124	/* 1) run (and print duration) */
125	pr_debug("calling  %lli_%pS @ %i\n", (long long)entry->cookie,
126		 entry->func, task_pid_nr(current));
127	calltime = ktime_get();
128
 
 
 
 
 
 
 
129	entry->func(entry->data, entry->cookie);
 
 
 
 
 
 
 
 
130
131	pr_debug("initcall %lli_%pS returned after %lld usecs\n",
132		 (long long)entry->cookie, entry->func,
133		 microseconds_since(calltime));
134
135	/* 2) remove self from the pending queues */
136	spin_lock_irqsave(&async_lock, flags);
137	list_del_init(&entry->domain_list);
138	list_del_init(&entry->global_list);
139
140	/* 3) free the entry */
141	kfree(entry);
142	atomic_dec(&entry_count);
143
144	spin_unlock_irqrestore(&async_lock, flags);
145
146	/* 4) wake up any waiters */
147	wake_up(&async_done);
148}
149
150static async_cookie_t __async_schedule_node_domain(async_func_t func,
151						   void *data, int node,
152						   struct async_domain *domain,
153						   struct async_entry *entry)
154{
155	async_cookie_t newcookie;
156	unsigned long flags;
157
158	INIT_LIST_HEAD(&entry->domain_list);
159	INIT_LIST_HEAD(&entry->global_list);
160	INIT_WORK(&entry->work, async_run_entry_fn);
161	entry->func = func;
162	entry->data = data;
163	entry->domain = domain;
164
165	spin_lock_irqsave(&async_lock, flags);
166
167	/* allocate cookie and queue */
168	newcookie = entry->cookie = next_cookie++;
169
170	list_add_tail(&entry->domain_list, &domain->pending);
171	if (domain->registered)
172		list_add_tail(&entry->global_list, &async_global_pending);
173
174	atomic_inc(&entry_count);
175	spin_unlock_irqrestore(&async_lock, flags);
176
177	/* schedule for execution */
178	queue_work_node(node, async_wq, &entry->work);
179
180	return newcookie;
181}
182
183/**
184 * async_schedule_node_domain - NUMA specific version of async_schedule_domain
185 * @func: function to execute asynchronously
186 * @data: data pointer to pass to the function
187 * @node: NUMA node that we want to schedule this on or close to
188 * @domain: the domain
189 *
190 * Returns an async_cookie_t that may be used for checkpointing later.
191 * @domain may be used in the async_synchronize_*_domain() functions to
192 * wait within a certain synchronization domain rather than globally.
193 *
194 * Note: This function may be called from atomic or non-atomic contexts.
195 *
196 * The node requested will be honored on a best effort basis. If the node
197 * has no CPUs associated with it then the work is distributed among all
198 * available CPUs.
199 */
200async_cookie_t async_schedule_node_domain(async_func_t func, void *data,
201					  int node, struct async_domain *domain)
202{
203	struct async_entry *entry;
204	unsigned long flags;
205	async_cookie_t newcookie;
206
207	/* allow irq-off callers */
208	entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
209
210	/*
211	 * If we're out of memory or if there's too much work
212	 * pending already, we execute synchronously.
213	 */
214	if (!entry || atomic_read(&entry_count) > MAX_WORK) {
215		kfree(entry);
216		spin_lock_irqsave(&async_lock, flags);
217		newcookie = next_cookie++;
218		spin_unlock_irqrestore(&async_lock, flags);
219
220		/* low on memory.. run synchronously */
221		func(data, newcookie);
222		return newcookie;
223	}
 
 
 
 
 
 
 
 
 
 
 
 
 
224
225	return __async_schedule_node_domain(func, data, node, domain, entry);
226}
227EXPORT_SYMBOL_GPL(async_schedule_node_domain);
228
229/**
230 * async_schedule_node - NUMA specific version of async_schedule
231 * @func: function to execute asynchronously
232 * @data: data pointer to pass to the function
233 * @node: NUMA node that we want to schedule this on or close to
234 *
235 * Returns an async_cookie_t that may be used for checkpointing later.
236 * Note: This function may be called from atomic or non-atomic contexts.
237 *
238 * The node requested will be honored on a best effort basis. If the node
239 * has no CPUs associated with it then the work is distributed among all
240 * available CPUs.
241 */
242async_cookie_t async_schedule_node(async_func_t func, void *data, int node)
243{
244	return async_schedule_node_domain(func, data, node, &async_dfl_domain);
245}
246EXPORT_SYMBOL_GPL(async_schedule_node);
247
248/**
249 * async_schedule_dev_nocall - A simplified variant of async_schedule_dev()
250 * @func: function to execute asynchronously
251 * @dev: device argument to be passed to function
 
252 *
253 * @dev is used as both the argument for the function and to provide NUMA
254 * context for where to run the function.
255 *
256 * If the asynchronous execution of @func is scheduled successfully, return
257 * true. Otherwise, do nothing and return false, unlike async_schedule_dev()
258 * that will run the function synchronously then.
259 */
260bool async_schedule_dev_nocall(async_func_t func, struct device *dev)
 
261{
262	struct async_entry *entry;
263
264	entry = kzalloc(sizeof(struct async_entry), GFP_KERNEL);
265
266	/* Give up if there is no memory or too much work. */
267	if (!entry || atomic_read(&entry_count) > MAX_WORK) {
268		kfree(entry);
269		return false;
270	}
271
272	__async_schedule_node_domain(func, dev, dev_to_node(dev),
273				     &async_dfl_domain, entry);
274	return true;
275}
 
276
277/**
278 * async_synchronize_full - synchronize all asynchronous function calls
279 *
280 * This function waits until all asynchronous function calls have been done.
281 */
282void async_synchronize_full(void)
283{
284	async_synchronize_full_domain(NULL);
 
 
285}
286EXPORT_SYMBOL_GPL(async_synchronize_full);
287
288/**
289 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
290 * @domain: the domain to synchronize
291 *
292 * This function waits until all asynchronous function calls for the
293 * synchronization domain specified by @domain have been done.
294 */
295void async_synchronize_full_domain(struct async_domain *domain)
296{
297	async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain);
298}
299EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
300
301/**
302 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
303 * @cookie: async_cookie_t to use as checkpoint
304 * @domain: the domain to synchronize (%NULL for all registered domains)
305 *
306 * This function waits until all asynchronous function calls for the
307 * synchronization domain specified by @domain submitted prior to @cookie
308 * have been done.
309 */
310void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *domain)
 
311{
312	ktime_t starttime;
313
314	pr_debug("async_waiting @ %i\n", task_pid_nr(current));
315	starttime = ktime_get();
 
 
316
317	wait_event(async_done, lowest_in_progress(domain) >= cookie);
318
319	pr_debug("async_continuing @ %i after %lli usec\n", task_pid_nr(current),
320		 microseconds_since(starttime));
 
 
 
 
 
 
321}
322EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
323
324/**
325 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
326 * @cookie: async_cookie_t to use as checkpoint
327 *
328 * This function waits until all asynchronous function calls prior to @cookie
329 * have been done.
330 */
331void async_synchronize_cookie(async_cookie_t cookie)
332{
333	async_synchronize_cookie_domain(cookie, &async_dfl_domain);
334}
335EXPORT_SYMBOL_GPL(async_synchronize_cookie);
336
337/**
338 * current_is_async - is %current an async worker task?
339 *
340 * Returns %true if %current is an async worker task.
341 */
342bool current_is_async(void)
343{
344	struct worker *worker = current_wq_worker();
345
346	return worker && worker->current_func == async_run_entry_fn;
347}
348EXPORT_SYMBOL_GPL(current_is_async);
349
350void __init async_init(void)
351{
352	/*
353	 * Async can schedule a number of interdependent work items. However,
354	 * unbound workqueues can handle only upto min_active interdependent
355	 * work items. The default min_active of 8 isn't sufficient for async
356	 * and can lead to stalls. Let's use a dedicated workqueue with raised
357	 * min_active.
358	 */
359	async_wq = alloc_workqueue("async", WQ_UNBOUND, 0);
360	BUG_ON(!async_wq);
361	workqueue_set_min_active(async_wq, WQ_DFL_ACTIVE);
362}