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 | |
12 | Goals and Theory of Operation |
13 | |
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. |
17 | |
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) |
23 | |
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). |
27 | |
28 | The async core will assign each scheduled event such a sequence cookie and |
29 | pass this to the called functions. |
30 | |
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. |
37 | |
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. |
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 | |
58 | static async_cookie_t next_cookie = 1; |
59 | |
60 | #define MAX_WORK 32768 |
61 | #define ASYNC_COOKIE_MAX ULLONG_MAX /* infinity cookie */ |
62 | |
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); |
66 | |
67 | struct 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 | |
77 | static DECLARE_WAIT_QUEUE_HEAD(async_done); |
78 | |
79 | static atomic_t entry_count; |
80 | |
81 | static 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 | |
87 | static 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(head: &domain->pending)) |
97 | first = list_first_entry(&domain->pending, |
98 | struct async_entry, domain_list); |
99 | } else { |
100 | if (!list_empty(head: &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(lock: &async_lock, flags); |
109 | return ret; |
110 | } |
111 | |
112 | /* |
113 | * pick the first pending entry and run it |
114 | */ |
115 | static 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: &entry->domain_list); |
136 | list_del_init(entry: &entry->global_list); |
137 | |
138 | /* 3) free the entry */ |
139 | kfree(objp: entry); |
140 | atomic_dec(v: &entry_count); |
141 | |
142 | spin_unlock_irqrestore(lock: &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 | */ |
165 | async_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(size: 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(v: &entry_count) > MAX_WORK) { |
180 | kfree(objp: entry); |
181 | spin_lock_irqsave(&async_lock, flags); |
182 | newcookie = next_cookie++; |
183 | spin_unlock_irqrestore(lock: &async_lock, flags); |
184 | |
185 | /* low on memory.. run synchronously */ |
186 | func(data, newcookie); |
187 | return newcookie; |
188 | } |
189 | INIT_LIST_HEAD(list: &entry->domain_list); |
190 | INIT_LIST_HEAD(list: &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(new: &entry->domain_list, head: &domain->pending); |
202 | if (domain->registered) |
203 | list_add_tail(new: &entry->global_list, head: &async_global_pending); |
204 | |
205 | atomic_inc(v: &entry_count); |
206 | spin_unlock_irqrestore(lock: &async_lock, flags); |
207 | |
208 | /* schedule for execution */ |
209 | queue_work_node(node, wq: system_unbound_wq, work: &entry->work); |
210 | |
211 | return newcookie; |
212 | } |
213 | EXPORT_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 | */ |
228 | async_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 | } |
232 | EXPORT_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 | */ |
239 | void async_synchronize_full(void) |
240 | { |
241 | async_synchronize_full_domain(NULL); |
242 | } |
243 | EXPORT_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 | */ |
252 | void async_synchronize_full_domain(struct async_domain *domain) |
253 | { |
254 | async_synchronize_cookie_domain(ASYNC_COOKIE_MAX, domain); |
255 | } |
256 | EXPORT_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 | */ |
267 | void 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 | } |
279 | EXPORT_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 | */ |
288 | void async_synchronize_cookie(async_cookie_t cookie) |
289 | { |
290 | async_synchronize_cookie_domain(cookie, &async_dfl_domain); |
291 | } |
292 | EXPORT_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 | */ |
299 | bool current_is_async(void) |
300 | { |
301 | struct worker *worker = current_wq_worker(); |
302 | |
303 | return worker && worker->current_func == async_run_entry_fn; |
304 | } |
305 | EXPORT_SYMBOL_GPL(current_is_async); |
306 | |