Loading...
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/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>
55
56#include "workqueue_internal.h"
57
58static async_cookie_t next_cookie = 1;
59
60#define MAX_WORK 32768
61#define ASYNC_COOKIE_MAX ULLONG_MAX /* infinity cookie */
62
63static LIST_HEAD(async_global_pending); /* pending from all registered doms */
64static ASYNC_DOMAIN(async_dfl_domain);
65static DEFINE_SPINLOCK(async_lock);
66
67struct async_entry {
68 struct list_head domain_list;
69 struct list_head global_list;
70 struct work_struct work;
71 async_cookie_t cookie;
72 async_func_t func;
73 void *data;
74 struct async_domain *domain;
75};
76
77static DECLARE_WAIT_QUEUE_HEAD(async_done);
78
79static atomic_t entry_count;
80
81static async_cookie_t lowest_in_progress(struct async_domain *domain)
82{
83 struct async_entry *first = NULL;
84 async_cookie_t ret = ASYNC_COOKIE_MAX;
85 unsigned long flags;
86
87 spin_lock_irqsave(&async_lock, flags);
88
89 if (domain) {
90 if (!list_empty(&domain->pending))
91 first = list_first_entry(&domain->pending,
92 struct async_entry, domain_list);
93 } else {
94 if (!list_empty(&async_global_pending))
95 first = list_first_entry(&async_global_pending,
96 struct async_entry, global_list);
97 }
98
99 if (first)
100 ret = first->cookie;
101
102 spin_unlock_irqrestore(&async_lock, flags);
103 return ret;
104}
105
106/*
107 * pick the first pending entry and run it
108 */
109static void async_run_entry_fn(struct work_struct *work)
110{
111 struct async_entry *entry =
112 container_of(work, struct async_entry, work);
113 unsigned long flags;
114 ktime_t uninitialized_var(calltime), delta, rettime;
115
116 /* 1) run (and print duration) */
117 if (initcall_debug && system_state < SYSTEM_RUNNING) {
118 pr_debug("calling %lli_%pS @ %i\n",
119 (long long)entry->cookie,
120 entry->func, task_pid_nr(current));
121 calltime = ktime_get();
122 }
123 entry->func(entry->data, entry->cookie);
124 if (initcall_debug && system_state < SYSTEM_RUNNING) {
125 rettime = ktime_get();
126 delta = ktime_sub(rettime, calltime);
127 pr_debug("initcall %lli_%pS returned 0 after %lld usecs\n",
128 (long long)entry->cookie,
129 entry->func,
130 (long long)ktime_to_ns(delta) >> 10);
131 }
132
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);
137
138 /* 3) free the entry */
139 kfree(entry);
140 atomic_dec(&entry_count);
141
142 spin_unlock_irqrestore(&async_lock, flags);
143
144 /* 4) wake up any waiters */
145 wake_up(&async_done);
146}
147
148/**
149 * async_schedule_node_domain - NUMA specific version of async_schedule_domain
150 * @func: function to execute asynchronously
151 * @data: data pointer to pass to the function
152 * @node: NUMA node that we want to schedule this on or close to
153 * @domain: the domain
154 *
155 * Returns an async_cookie_t that may be used for checkpointing later.
156 * @domain may be used in the async_synchronize_*_domain() functions to
157 * wait within a certain synchronization domain rather than globally.
158 *
159 * Note: This function may be called from atomic or non-atomic contexts.
160 *
161 * The node requested will be honored on a best effort basis. If the node
162 * has no CPUs associated with it then the work is distributed among all
163 * available CPUs.
164 */
165async_cookie_t async_schedule_node_domain(async_func_t func, void *data,
166 int node, struct async_domain *domain)
167{
168 struct async_entry *entry;
169 unsigned long flags;
170 async_cookie_t newcookie;
171
172 /* allow irq-off callers */
173 entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
174
175 /*
176 * If we're out of memory or if there's too much work
177 * pending already, we execute synchronously.
178 */
179 if (!entry || atomic_read(&entry_count) > MAX_WORK) {
180 kfree(entry);
181 spin_lock_irqsave(&async_lock, flags);
182 newcookie = next_cookie++;
183 spin_unlock_irqrestore(&async_lock, flags);
184
185 /* low on memory.. run synchronously */
186 func(data, newcookie);
187 return newcookie;
188 }
189 INIT_LIST_HEAD(&entry->domain_list);
190 INIT_LIST_HEAD(&entry->global_list);
191 INIT_WORK(&entry->work, async_run_entry_fn);
192 entry->func = func;
193 entry->data = data;
194 entry->domain = domain;
195
196 spin_lock_irqsave(&async_lock, flags);
197
198 /* allocate cookie and queue */
199 newcookie = entry->cookie = next_cookie++;
200
201 list_add_tail(&entry->domain_list, &domain->pending);
202 if (domain->registered)
203 list_add_tail(&entry->global_list, &async_global_pending);
204
205 atomic_inc(&entry_count);
206 spin_unlock_irqrestore(&async_lock, flags);
207
208 /* mark that this task has queued an async job, used by module init */
209 current->flags |= PF_USED_ASYNC;
210
211 /* schedule for execution */
212 queue_work_node(node, system_unbound_wq, &entry->work);
213
214 return newcookie;
215}
216EXPORT_SYMBOL_GPL(async_schedule_node_domain);
217
218/**
219 * async_schedule_node - NUMA specific version of async_schedule
220 * @func: function to execute asynchronously
221 * @data: data pointer to pass to the function
222 * @node: NUMA node that we want to schedule this on or close to
223 *
224 * Returns an async_cookie_t that may be used for checkpointing later.
225 * Note: This function may be called from atomic or non-atomic contexts.
226 *
227 * The node requested will be honored on a best effort basis. If the node
228 * has no CPUs associated with it then the work is distributed among all
229 * available CPUs.
230 */
231async_cookie_t async_schedule_node(async_func_t func, void *data, int node)
232{
233 return async_schedule_node_domain(func, data, node, &async_dfl_domain);
234}
235EXPORT_SYMBOL_GPL(async_schedule_node);
236
237/**
238 * async_synchronize_full - synchronize all asynchronous function calls
239 *
240 * This function waits until all asynchronous function calls have been done.
241 */
242void async_synchronize_full(void)
243{
244 async_synchronize_full_domain(NULL);
245}
246EXPORT_SYMBOL_GPL(async_synchronize_full);
247
248/**
249 * async_unregister_domain - ensure no more anonymous waiters on this domain
250 * @domain: idle domain to flush out of any async_synchronize_full instances
251 *
252 * async_synchronize_{cookie|full}_domain() are not flushed since callers
253 * of these routines should know the lifetime of @domain
254 *
255 * Prefer ASYNC_DOMAIN_EXCLUSIVE() declarations over flushing
256 */
257void async_unregister_domain(struct async_domain *domain)
258{
259 spin_lock_irq(&async_lock);
260 WARN_ON(!domain->registered || !list_empty(&domain->pending));
261 domain->registered = 0;
262 spin_unlock_irq(&async_lock);
263}
264EXPORT_SYMBOL_GPL(async_unregister_domain);
265
266/**
267 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
268 * @domain: the domain to synchronize
269 *
270 * This function waits until all asynchronous function calls for the
271 * synchronization domain specified by @domain have been done.
272 */
273void async_synchronize_full_domain(struct async_domain *domain)
274{
275 async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain);
276}
277EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
278
279/**
280 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
281 * @cookie: async_cookie_t to use as checkpoint
282 * @domain: the domain to synchronize (%NULL for all registered domains)
283 *
284 * This function waits until all asynchronous function calls for the
285 * synchronization domain specified by @domain submitted prior to @cookie
286 * have been done.
287 */
288void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *domain)
289{
290 ktime_t uninitialized_var(starttime), delta, endtime;
291
292 if (initcall_debug && system_state < SYSTEM_RUNNING) {
293 pr_debug("async_waiting @ %i\n", task_pid_nr(current));
294 starttime = ktime_get();
295 }
296
297 wait_event(async_done, lowest_in_progress(domain) >= cookie);
298
299 if (initcall_debug && system_state < SYSTEM_RUNNING) {
300 endtime = ktime_get();
301 delta = ktime_sub(endtime, starttime);
302
303 pr_debug("async_continuing @ %i after %lli usec\n",
304 task_pid_nr(current),
305 (long long)ktime_to_ns(delta) >> 10);
306 }
307}
308EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
309
310/**
311 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
312 * @cookie: async_cookie_t to use as checkpoint
313 *
314 * This function waits until all asynchronous function calls prior to @cookie
315 * have been done.
316 */
317void async_synchronize_cookie(async_cookie_t cookie)
318{
319 async_synchronize_cookie_domain(cookie, &async_dfl_domain);
320}
321EXPORT_SYMBOL_GPL(async_synchronize_cookie);
322
323/**
324 * current_is_async - is %current an async worker task?
325 *
326 * Returns %true if %current is an async worker task.
327 */
328bool current_is_async(void)
329{
330 struct worker *worker = current_wq_worker();
331
332 return worker && worker->current_func == async_run_entry_fn;
333}
334EXPORT_SYMBOL_GPL(current_is_async);
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/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
62static async_cookie_t next_cookie = 1;
63
64#define MAX_WORK 32768
65#define ASYNC_COOKIE_MAX ULLONG_MAX /* infinity cookie */
66
67static LIST_HEAD(async_global_pending); /* pending from all registered doms */
68static ASYNC_DOMAIN(async_dfl_domain);
69static DEFINE_SPINLOCK(async_lock);
70
71struct 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
81static DECLARE_WAIT_QUEUE_HEAD(async_done);
82
83static atomic_t entry_count;
84
85static 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 */
113static 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
152static 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 */
211async_cookie_t async_schedule(async_func_t func, void *data)
212{
213 return __async_schedule(func, data, &async_dfl_domain);
214}
215EXPORT_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 */
229async_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}
234EXPORT_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 */
241void async_synchronize_full(void)
242{
243 async_synchronize_full_domain(NULL);
244}
245EXPORT_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 */
256void 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}
263EXPORT_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 */
272void async_synchronize_full_domain(struct async_domain *domain)
273{
274 async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain);
275}
276EXPORT_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 */
287void 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}
307EXPORT_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 */
316void async_synchronize_cookie(async_cookie_t cookie)
317{
318 async_synchronize_cookie_domain(cookie, &async_dfl_domain);
319}
320EXPORT_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 */
327bool current_is_async(void)
328{
329 struct worker *worker = current_wq_worker();
330
331 return worker && worker->current_func == async_run_entry_fn;
332}
333EXPORT_SYMBOL_GPL(current_is_async);