1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
use fmt;
use UnsafeCell;
use AtomicUsize;
use ;
use Waker;
/// A synchronization primitive for task wakeup.
///
/// Sometimes the task interested in a given event will change over time.
/// An `AtomicWaker` can coordinate concurrent notifications with the consumer
/// potentially "updating" the underlying task to wake up. This is useful in
/// scenarios where a computation completes in another thread and wants to
/// notify the consumer, but the consumer is in the process of being migrated to
/// a new logical task.
///
/// Consumers should call `register` before checking the result of a computation
/// and producers should call `wake` after producing the computation (this
/// differs from the usual `thread::park` pattern). It is also permitted for
/// `wake` to be called **before** `register`. This results in a no-op.
///
/// A single `AtomicWaker` may be reused for any number of calls to `register` or
/// `wake`.
///
/// `AtomicWaker` does not provide any memory ordering guarantees, as such the
/// user should use caution and use other synchronization primitives to guard
/// the result of the underlying computation.
// `AtomicWaker` is a multi-consumer, single-producer transfer cell. The cell
// stores a `Waker` value produced by calls to `register` and many threads can
// race to take the waker (to wake it) by calling `wake`.
//
// If a new `Waker` instance is produced by calling `register` before an
// existing one is consumed, then the existing one is overwritten.
//
// While `AtomicWaker` is single-producer, the implementation ensures memory
// safety. In the event of concurrent calls to `register`, there will be a
// single winner whose waker will get stored in the cell. The losers will not
// have their tasks woken. As such, callers should ensure to add synchronization
// to calls to `register`.
//
// The implementation uses a single `AtomicUsize` value to coordinate access to
// the `Waker` cell. There are two bits that are operated on independently.
// These are represented by `REGISTERING` and `WAKING`.
//
// The `REGISTERING` bit is set when a producer enters the critical section. The
// `WAKING` bit is set when a consumer enters the critical section. Neither bit
// being set is represented by `WAITING`.
//
// A thread obtains an exclusive lock on the waker cell by transitioning the
// state from `WAITING` to `REGISTERING` or `WAKING`, depending on the operation
// the thread wishes to perform. When this transition is made, it is guaranteed
// that no other thread will access the waker cell.
//
// # Registering
//
// On a call to `register`, an attempt to transition the state from WAITING to
// REGISTERING is made. On success, the caller obtains a lock on the waker cell.
//
// If the lock is obtained, then the thread sets the waker cell to the waker
// provided as an argument. Then it attempts to transition the state back from
// `REGISTERING` -> `WAITING`.
//
// If this transition is successful, then the registering process is complete
// and the next call to `wake` will observe the waker.
//
// If the transition fails, then there was a concurrent call to `wake` that was
// unable to access the waker cell (due to the registering thread holding the
// lock). To handle this, the registering thread removes the waker it just set
// from the cell and calls `wake` on it. This call to wake represents the
// attempt to wake by the other thread (that set the `WAKING` bit). The state is
// then transitioned from `REGISTERING | WAKING` back to `WAITING`. This
// transition must succeed because, at this point, the state cannot be
// transitioned by another thread.
//
// # Waking
//
// On a call to `wake`, an attempt to transition the state from `WAITING` to
// `WAKING` is made. On success, the caller obtains a lock on the waker cell.
//
// If the lock is obtained, then the thread takes ownership of the current value
// in the waker cell, and calls `wake` on it. The state is then transitioned
// back to `WAITING`. This transition must succeed as, at this point, the state
// cannot be transitioned by another thread.
//
// If the thread is unable to obtain the lock, the `WAKING` bit is still. This
// is because it has either been set by the current thread but the previous
// value included the `REGISTERING` bit **or** a concurrent thread is in the
// `WAKING` critical section. Either way, no action must be taken.
//
// If the current thread is the only concurrent call to `wake` and another
// thread is in the `register` critical section, when the other thread **exits**
// the `register` critical section, it will observe the `WAKING` bit and handle
// the wake itself.
//
// If another thread is in the `wake` critical section, then it will handle
// waking the task.
//
// # A potential race (is safely handled).
//
// Imagine the following situation:
//
// * Thread A obtains the `wake` lock and wakes a task.
//
// * Before thread A releases the `wake` lock, the woken task is scheduled.
//
// * Thread B attempts to wake the task. In theory this should result in the
// task being woken, but it cannot because thread A still holds the wake lock.
//
// This case is handled by requiring users of `AtomicWaker` to call `register`
// **before** attempting to observe the application state change that resulted
// in the task being awoken. The wakers also change the application state before
// calling wake.
//
// Because of this, the waker will do one of two things.
//
// 1) Observe the application state change that Thread B is woken for. In this
// case, it is OK for Thread B's wake to be lost.
//
// 2) Call register before attempting to observe the application state. Since
// Thread A still holds the `wake` lock, the call to `register` will result
// in the task waking itself and get scheduled again.
/// Idle state
const WAITING: usize = 0;
/// A new waker value is being registered with the `AtomicWaker` cell.
const REGISTERING: usize = 0b01;
/// The waker currently registered with the `AtomicWaker` cell is being woken.
const WAKING: usize = 0b10;
unsafe
unsafe