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 long long microseconds_since(ktime_t start)
82{
83 ktime_t now = ktime_get();
84 return ktime_to_ns(ktime_sub(now, start)) >> 10;
85}
86
87static async_cookie_t lowest_in_progress(struct async_domain *domain)
88{
89 struct async_entry *first = NULL;
90 async_cookie_t ret = ASYNC_COOKIE_MAX;
91 unsigned long flags;
92
93 spin_lock_irqsave(&async_lock, flags);
94
95 if (domain) {
96 if (!list_empty(&domain->pending))
97 first = list_first_entry(&domain->pending,
98 struct async_entry, domain_list);
99 } else {
100 if (!list_empty(&async_global_pending))
101 first = list_first_entry(&async_global_pending,
102 struct async_entry, global_list);
103 }
104
105 if (first)
106 ret = first->cookie;
107
108 spin_unlock_irqrestore(&async_lock, flags);
109 return ret;
110}
111
112/*
113 * pick the first pending entry and run it
114 */
115static void async_run_entry_fn(struct work_struct *work)
116{
117 struct async_entry *entry =
118 container_of(work, struct async_entry, work);
119 unsigned long flags;
120 ktime_t calltime;
121
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();
126
127 entry->func(entry->data, entry->cookie);
128
129 pr_debug("initcall %lli_%pS returned after %lld usecs\n",
130 (long long)entry->cookie, entry->func,
131 microseconds_since(calltime));
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 /* schedule for execution */
209 queue_work_node(node, system_unbound_wq, &entry->work);
210
211 return newcookie;
212}
213EXPORT_SYMBOL_GPL(async_schedule_node_domain);
214
215/**
216 * async_schedule_node - NUMA specific version of async_schedule
217 * @func: function to execute asynchronously
218 * @data: data pointer to pass to the function
219 * @node: NUMA node that we want to schedule this on or close to
220 *
221 * Returns an async_cookie_t that may be used for checkpointing later.
222 * Note: This function may be called from atomic or non-atomic contexts.
223 *
224 * The node requested will be honored on a best effort basis. If the node
225 * has no CPUs associated with it then the work is distributed among all
226 * available CPUs.
227 */
228async_cookie_t async_schedule_node(async_func_t func, void *data, int node)
229{
230 return async_schedule_node_domain(func, data, node, &async_dfl_domain);
231}
232EXPORT_SYMBOL_GPL(async_schedule_node);
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 async_synchronize_full_domain(NULL);
242}
243EXPORT_SYMBOL_GPL(async_synchronize_full);
244
245/**
246 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
247 * @domain: the domain to synchronize
248 *
249 * This function waits until all asynchronous function calls for the
250 * synchronization domain specified by @domain have been done.
251 */
252void async_synchronize_full_domain(struct async_domain *domain)
253{
254 async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain);
255}
256EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
257
258/**
259 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
260 * @cookie: async_cookie_t to use as checkpoint
261 * @domain: the domain to synchronize (%NULL for all registered domains)
262 *
263 * This function waits until all asynchronous function calls for the
264 * synchronization domain specified by @domain submitted prior to @cookie
265 * have been done.
266 */
267void async_synchronize_cookie_domain(async_cookie_t cookie, struct async_domain *domain)
268{
269 ktime_t starttime;
270
271 pr_debug("async_waiting @ %i\n", task_pid_nr(current));
272 starttime = ktime_get();
273
274 wait_event(async_done, lowest_in_progress(domain) >= cookie);
275
276 pr_debug("async_continuing @ %i after %lli usec\n", task_pid_nr(current),
277 microseconds_since(starttime));
278}
279EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
280
281/**
282 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
283 * @cookie: async_cookie_t to use as checkpoint
284 *
285 * This function waits until all asynchronous function calls prior to @cookie
286 * have been done.
287 */
288void async_synchronize_cookie(async_cookie_t cookie)
289{
290 async_synchronize_cookie_domain(cookie, &async_dfl_domain);
291}
292EXPORT_SYMBOL_GPL(async_synchronize_cookie);
293
294/**
295 * current_is_async - is %current an async worker task?
296 *
297 * Returns %true if %current is an async worker task.
298 */
299bool current_is_async(void)
300{
301 struct worker *worker = current_wq_worker();
302
303 return worker && worker->current_func == async_run_entry_fn;
304}
305EXPORT_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
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
81
82/*
83 * MUST be called with the lock held!
84 */
85static async_cookie_t __lowest_in_progress(struct list_head *running)
86{
87 struct async_entry *entry;
88
89 if (!running) { /* just check the entry count */
90 if (atomic_read(&entry_count))
91 return 0; /* smaller than any cookie */
92 else
93 return next_cookie;
94 }
95
96 if (!list_empty(running)) {
97 entry = list_first_entry(running,
98 struct async_entry, list);
99 return entry->cookie;
100 }
101
102 list_for_each_entry(entry, &async_pending, list)
103 if (entry->running == running)
104 return entry->cookie;
105
106 return next_cookie; /* "infinity" value */
107}
108
109static async_cookie_t lowest_in_progress(struct list_head *running)
110{
111 unsigned long flags;
112 async_cookie_t ret;
113
114 spin_lock_irqsave(&async_lock, flags);
115 ret = __lowest_in_progress(running);
116 spin_unlock_irqrestore(&async_lock, flags);
117 return ret;
118}
119
120/*
121 * pick the first pending entry and run it
122 */
123static void async_run_entry_fn(struct work_struct *work)
124{
125 struct async_entry *entry =
126 container_of(work, struct async_entry, work);
127 unsigned long flags;
128 ktime_t uninitialized_var(calltime), delta, rettime;
129
130 /* 1) move self to the running queue */
131 spin_lock_irqsave(&async_lock, flags);
132 list_move_tail(&entry->list, entry->running);
133 spin_unlock_irqrestore(&async_lock, flags);
134
135 /* 2) run (and print duration) */
136 if (initcall_debug && system_state == SYSTEM_BOOTING) {
137 printk(KERN_DEBUG "calling %lli_%pF @ %i\n",
138 (long long)entry->cookie,
139 entry->func, task_pid_nr(current));
140 calltime = ktime_get();
141 }
142 entry->func(entry->data, entry->cookie);
143 if (initcall_debug && system_state == SYSTEM_BOOTING) {
144 rettime = ktime_get();
145 delta = ktime_sub(rettime, calltime);
146 printk(KERN_DEBUG "initcall %lli_%pF returned 0 after %lld usecs\n",
147 (long long)entry->cookie,
148 entry->func,
149 (long long)ktime_to_ns(delta) >> 10);
150 }
151
152 /* 3) remove self from the running queue */
153 spin_lock_irqsave(&async_lock, flags);
154 list_del(&entry->list);
155
156 /* 4) free the entry */
157 kfree(entry);
158 atomic_dec(&entry_count);
159
160 spin_unlock_irqrestore(&async_lock, flags);
161
162 /* 5) wake up any waiters */
163 wake_up(&async_done);
164}
165
166static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct list_head *running)
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 ptr(data, newcookie);
187 return newcookie;
188 }
189 INIT_WORK(&entry->work, async_run_entry_fn);
190 entry->func = ptr;
191 entry->data = data;
192 entry->running = running;
193
194 spin_lock_irqsave(&async_lock, flags);
195 newcookie = entry->cookie = next_cookie++;
196 list_add_tail(&entry->list, &async_pending);
197 atomic_inc(&entry_count);
198 spin_unlock_irqrestore(&async_lock, flags);
199
200 /* schedule for execution */
201 queue_work(system_unbound_wq, &entry->work);
202
203 return newcookie;
204}
205
206/**
207 * async_schedule - schedule a function for asynchronous execution
208 * @ptr: function to execute asynchronously
209 * @data: data pointer to pass to the function
210 *
211 * Returns an async_cookie_t that may be used for checkpointing later.
212 * Note: This function may be called from atomic or non-atomic contexts.
213 */
214async_cookie_t async_schedule(async_func_ptr *ptr, void *data)
215{
216 return __async_schedule(ptr, data, &async_running);
217}
218EXPORT_SYMBOL_GPL(async_schedule);
219
220/**
221 * async_schedule_domain - schedule a function for asynchronous execution within a certain domain
222 * @ptr: function to execute asynchronously
223 * @data: data pointer to pass to the function
224 * @running: running list for the domain
225 *
226 * Returns an async_cookie_t that may be used for checkpointing later.
227 * @running may be used in the async_synchronize_*_domain() functions
228 * to wait within a certain synchronization domain rather than globally.
229 * A synchronization domain is specified via the running queue @running to use.
230 * Note: This function may be called from atomic or non-atomic contexts.
231 */
232async_cookie_t async_schedule_domain(async_func_ptr *ptr, void *data,
233 struct list_head *running)
234{
235 return __async_schedule(ptr, data, running);
236}
237EXPORT_SYMBOL_GPL(async_schedule_domain);
238
239/**
240 * async_synchronize_full - synchronize all asynchronous function calls
241 *
242 * This function waits until all asynchronous function calls have been done.
243 */
244void async_synchronize_full(void)
245{
246 async_synchronize_cookie_domain(next_cookie, NULL);
247}
248EXPORT_SYMBOL_GPL(async_synchronize_full);
249
250/**
251 * async_synchronize_full_domain - synchronize all asynchronous function within a certain domain
252 * @list: running list to synchronize on
253 *
254 * This function waits until all asynchronous function calls for the
255 * synchronization domain specified by the running list @list have been done.
256 */
257void async_synchronize_full_domain(struct list_head *list)
258{
259 async_synchronize_cookie_domain(next_cookie, list);
260}
261EXPORT_SYMBOL_GPL(async_synchronize_full_domain);
262
263/**
264 * async_synchronize_cookie_domain - synchronize asynchronous function calls within a certain domain with cookie checkpointing
265 * @cookie: async_cookie_t to use as checkpoint
266 * @running: running list to synchronize on, NULL indicates all lists
267 *
268 * This function waits until all asynchronous function calls for the
269 * synchronization domain specified by the running list @list submitted
270 * prior to @cookie have been done.
271 */
272void async_synchronize_cookie_domain(async_cookie_t cookie,
273 struct list_head *running)
274{
275 ktime_t uninitialized_var(starttime), delta, endtime;
276
277 if (initcall_debug && system_state == SYSTEM_BOOTING) {
278 printk(KERN_DEBUG "async_waiting @ %i\n", task_pid_nr(current));
279 starttime = ktime_get();
280 }
281
282 wait_event(async_done, lowest_in_progress(running) >= cookie);
283
284 if (initcall_debug && system_state == SYSTEM_BOOTING) {
285 endtime = ktime_get();
286 delta = ktime_sub(endtime, starttime);
287
288 printk(KERN_DEBUG "async_continuing @ %i after %lli usec\n",
289 task_pid_nr(current),
290 (long long)ktime_to_ns(delta) >> 10);
291 }
292}
293EXPORT_SYMBOL_GPL(async_synchronize_cookie_domain);
294
295/**
296 * async_synchronize_cookie - synchronize asynchronous function calls with cookie checkpointing
297 * @cookie: async_cookie_t to use as checkpoint
298 *
299 * This function waits until all asynchronous function calls prior to @cookie
300 * have been done.
301 */
302void async_synchronize_cookie(async_cookie_t cookie)
303{
304 async_synchronize_cookie_domain(cookie, &async_running);
305}
306EXPORT_SYMBOL_GPL(async_synchronize_cookie);