1 | //! `futures::task::AtomicWaker` extracted into its own crate. |
2 | //! |
3 | //! # Features |
4 | //! |
5 | //! This crate adds a feature, `portable-atomic`, which uses a polyfill |
6 | //! from the [`portable-atomic`] crate in order to provide functionality |
7 | //! to targets without atomics. See the [`README`] for the [`portable-atomic`] |
8 | //! crate for more information on how to use it. |
9 | //! |
10 | //! [`portable-atomic`]: https://crates.io/crates/portable-atomic |
11 | //! [`README`]: https://github.com/taiki-e/portable-atomic/blob/main/README.md#optional-cfg |
12 | |
13 | #![no_std ] |
14 | #![doc ( |
15 | html_favicon_url = "https://raw.githubusercontent.com/smol-rs/smol/master/assets/images/logo_fullsize_transparent.png" |
16 | )] |
17 | #![doc ( |
18 | html_logo_url = "https://raw.githubusercontent.com/smol-rs/smol/master/assets/images/logo_fullsize_transparent.png" |
19 | )] |
20 | |
21 | use core::cell::UnsafeCell; |
22 | use core::fmt; |
23 | use core::sync::atomic::Ordering::{AcqRel, Acquire, Release}; |
24 | use core::task::Waker; |
25 | |
26 | #[cfg (not(feature = "portable-atomic" ))] |
27 | use core::sync::atomic::AtomicUsize; |
28 | #[cfg (feature = "portable-atomic" )] |
29 | use portable_atomic::AtomicUsize; |
30 | |
31 | /// A synchronization primitive for task wakeup. |
32 | /// |
33 | /// Sometimes the task interested in a given event will change over time. |
34 | /// An `AtomicWaker` can coordinate concurrent notifications with the consumer |
35 | /// potentially "updating" the underlying task to wake up. This is useful in |
36 | /// scenarios where a computation completes in another thread and wants to |
37 | /// notify the consumer, but the consumer is in the process of being migrated to |
38 | /// a new logical task. |
39 | /// |
40 | /// Consumers should call `register` before checking the result of a computation |
41 | /// and producers should call `wake` after producing the computation (this |
42 | /// differs from the usual `thread::park` pattern). It is also permitted for |
43 | /// `wake` to be called **before** `register`. This results in a no-op. |
44 | /// |
45 | /// A single `AtomicWaker` may be reused for any number of calls to `register` or |
46 | /// `wake`. |
47 | /// |
48 | /// # Memory ordering |
49 | /// |
50 | /// Calling `register` "acquires" all memory "released" by calls to `wake` |
51 | /// before the call to `register`. Later calls to `wake` will wake the |
52 | /// registered waker (on contention this wake might be triggered in `register`). |
53 | /// |
54 | /// For concurrent calls to `register` (should be avoided) the ordering is only |
55 | /// guaranteed for the winning call. |
56 | /// |
57 | /// # Examples |
58 | /// |
59 | /// Here is a simple example providing a `Flag` that can be signalled manually |
60 | /// when it is ready. |
61 | /// |
62 | /// ``` |
63 | /// use futures::future::Future; |
64 | /// use futures::task::{Context, Poll, AtomicWaker}; |
65 | /// use std::sync::Arc; |
66 | /// use std::sync::atomic::AtomicBool; |
67 | /// use std::sync::atomic::Ordering::Relaxed; |
68 | /// use std::pin::Pin; |
69 | /// |
70 | /// struct Inner { |
71 | /// waker: AtomicWaker, |
72 | /// set: AtomicBool, |
73 | /// } |
74 | /// |
75 | /// #[derive(Clone)] |
76 | /// struct Flag(Arc<Inner>); |
77 | /// |
78 | /// impl Flag { |
79 | /// pub fn new() -> Self { |
80 | /// Flag(Arc::new(Inner { |
81 | /// waker: AtomicWaker::new(), |
82 | /// set: AtomicBool::new(false), |
83 | /// })) |
84 | /// } |
85 | /// |
86 | /// pub fn signal(&self) { |
87 | /// self.0.set.store(true, Relaxed); |
88 | /// self.0.waker.wake(); |
89 | /// } |
90 | /// } |
91 | /// |
92 | /// impl Future for Flag { |
93 | /// type Output = (); |
94 | /// |
95 | /// fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<()> { |
96 | /// // quick check to avoid registration if already done. |
97 | /// if self.0.set.load(Relaxed) { |
98 | /// return Poll::Ready(()); |
99 | /// } |
100 | /// |
101 | /// self.0.waker.register(cx.waker()); |
102 | /// |
103 | /// // Need to check condition **after** `register` to avoid a race |
104 | /// // condition that would result in lost notifications. |
105 | /// if self.0.set.load(Relaxed) { |
106 | /// Poll::Ready(()) |
107 | /// } else { |
108 | /// Poll::Pending |
109 | /// } |
110 | /// } |
111 | /// } |
112 | /// ``` |
113 | pub struct AtomicWaker { |
114 | state: AtomicUsize, |
115 | waker: UnsafeCell<Option<Waker>>, |
116 | } |
117 | |
118 | // `AtomicWaker` is a multi-consumer, single-producer transfer cell. The cell |
119 | // stores a `Waker` value produced by calls to `register` and many threads can |
120 | // race to take the waker (to wake it) by calling `wake`. |
121 | // |
122 | // If a new `Waker` instance is produced by calling `register` before an |
123 | // existing one is consumed, then the existing one is overwritten. |
124 | // |
125 | // While `AtomicWaker` is single-producer, the implementation ensures memory |
126 | // safety. In the event of concurrent calls to `register`, there will be a |
127 | // single winner whose waker will get stored in the cell. The losers will not |
128 | // have their tasks woken. As such, callers should ensure to add synchronization |
129 | // to calls to `register`. |
130 | // |
131 | // The implementation uses a single `AtomicUsize` value to coordinate access to |
132 | // the `Waker` cell. There are two bits that are operated on independently. |
133 | // These are represented by `REGISTERING` and `WAKING`. |
134 | // |
135 | // The `REGISTERING` bit is set when a producer enters the critical section. The |
136 | // `WAKING` bit is set when a consumer enters the critical section. Neither bit |
137 | // being set is represented by `WAITING`. |
138 | // |
139 | // A thread obtains an exclusive lock on the waker cell by transitioning the |
140 | // state from `WAITING` to `REGISTERING` or `WAKING`, depending on the operation |
141 | // the thread wishes to perform. When this transition is made, it is guaranteed |
142 | // that no other thread will access the waker cell. |
143 | // |
144 | // # Registering |
145 | // |
146 | // On a call to `register`, an attempt to transition the state from WAITING to |
147 | // REGISTERING is made. On success, the caller obtains a lock on the waker cell. |
148 | // |
149 | // If the lock is obtained, then the thread sets the waker cell to the waker |
150 | // provided as an argument. Then it attempts to transition the state back from |
151 | // `REGISTERING` -> `WAITING`. |
152 | // |
153 | // If this transition is successful, then the registering process is complete |
154 | // and the next call to `wake` will observe the waker. |
155 | // |
156 | // If the transition fails, then there was a concurrent call to `wake` that was |
157 | // unable to access the waker cell (due to the registering thread holding the |
158 | // lock). To handle this, the registering thread removes the waker it just set |
159 | // from the cell and calls `wake` on it. This call to wake represents the |
160 | // attempt to wake by the other thread (that set the `WAKING` bit). The state is |
161 | // then transitioned from `REGISTERING | WAKING` back to `WAITING`. This |
162 | // transition must succeed because, at this point, the state cannot be |
163 | // transitioned by another thread. |
164 | // |
165 | // # Waking |
166 | // |
167 | // On a call to `wake`, an attempt to transition the state from `WAITING` to |
168 | // `WAKING` is made. On success, the caller obtains a lock on the waker cell. |
169 | // |
170 | // If the lock is obtained, then the thread takes ownership of the current value |
171 | // in the waker cell, and calls `wake` on it. The state is then transitioned |
172 | // back to `WAITING`. This transition must succeed as, at this point, the state |
173 | // cannot be transitioned by another thread. |
174 | // |
175 | // If the thread is unable to obtain the lock, the `WAKING` bit is still. This |
176 | // is because it has either been set by the current thread but the previous |
177 | // value included the `REGISTERING` bit **or** a concurrent thread is in the |
178 | // `WAKING` critical section. Either way, no action must be taken. |
179 | // |
180 | // If the current thread is the only concurrent call to `wake` and another |
181 | // thread is in the `register` critical section, when the other thread **exits** |
182 | // the `register` critical section, it will observe the `WAKING` bit and handle |
183 | // the wake itself. |
184 | // |
185 | // If another thread is in the `wake` critical section, then it will handle |
186 | // waking the task. |
187 | // |
188 | // # A potential race (is safely handled). |
189 | // |
190 | // Imagine the following situation: |
191 | // |
192 | // * Thread A obtains the `wake` lock and wakes a task. |
193 | // |
194 | // * Before thread A releases the `wake` lock, the woken task is scheduled. |
195 | // |
196 | // * Thread B attempts to wake the task. In theory this should result in the |
197 | // task being woken, but it cannot because thread A still holds the wake lock. |
198 | // |
199 | // This case is handled by requiring users of `AtomicWaker` to call `register` |
200 | // **before** attempting to observe the application state change that resulted |
201 | // in the task being awoken. The wakers also change the application state before |
202 | // calling wake. |
203 | // |
204 | // Because of this, the waker will do one of two things. |
205 | // |
206 | // 1) Observe the application state change that Thread B is woken for. In this |
207 | // case, it is OK for Thread B's wake to be lost. |
208 | // |
209 | // 2) Call register before attempting to observe the application state. Since |
210 | // Thread A still holds the `wake` lock, the call to `register` will result |
211 | // in the task waking itself and get scheduled again. |
212 | |
213 | /// Idle state |
214 | const WAITING: usize = 0; |
215 | |
216 | /// A new waker value is being registered with the `AtomicWaker` cell. |
217 | const REGISTERING: usize = 0b01; |
218 | |
219 | /// The waker currently registered with the `AtomicWaker` cell is being woken. |
220 | const WAKING: usize = 0b10; |
221 | |
222 | impl AtomicWaker { |
223 | /// Create an `AtomicWaker`. |
224 | pub const fn new() -> Self { |
225 | // Make sure that task is Sync |
226 | trait AssertSync: Sync {} |
227 | impl AssertSync for Waker {} |
228 | |
229 | AtomicWaker { |
230 | state: AtomicUsize::new(WAITING), |
231 | waker: UnsafeCell::new(None), |
232 | } |
233 | } |
234 | |
235 | /// Registers the waker to be notified on calls to `wake`. |
236 | /// |
237 | /// The new task will take place of any previous tasks that were registered |
238 | /// by previous calls to `register`. Any calls to `wake` that happen after |
239 | /// a call to `register` (as defined by the memory ordering rules), will |
240 | /// notify the `register` caller's task and deregister the waker from future |
241 | /// notifications. Because of this, callers should ensure `register` gets |
242 | /// invoked with a new `Waker` **each** time they require a wakeup. |
243 | /// |
244 | /// It is safe to call `register` with multiple other threads concurrently |
245 | /// calling `wake`. This will result in the `register` caller's current |
246 | /// task being notified once. |
247 | /// |
248 | /// This function is safe to call concurrently, but this is generally a bad |
249 | /// idea. Concurrent calls to `register` will attempt to register different |
250 | /// tasks to be notified. One of the callers will win and have its task set, |
251 | /// but there is no guarantee as to which caller will succeed. |
252 | /// |
253 | /// # Examples |
254 | /// |
255 | /// Here is how `register` is used when implementing a flag. |
256 | /// |
257 | /// ``` |
258 | /// use futures::future::Future; |
259 | /// use futures::task::{Context, Poll, AtomicWaker}; |
260 | /// use std::sync::atomic::AtomicBool; |
261 | /// use std::sync::atomic::Ordering::Relaxed; |
262 | /// use std::pin::Pin; |
263 | /// |
264 | /// struct Flag { |
265 | /// waker: AtomicWaker, |
266 | /// set: AtomicBool, |
267 | /// } |
268 | /// |
269 | /// impl Future for Flag { |
270 | /// type Output = (); |
271 | /// |
272 | /// fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<()> { |
273 | /// // Register **before** checking `set` to avoid a race condition |
274 | /// // that would result in lost notifications. |
275 | /// self.waker.register(cx.waker()); |
276 | /// |
277 | /// if self.set.load(Relaxed) { |
278 | /// Poll::Ready(()) |
279 | /// } else { |
280 | /// Poll::Pending |
281 | /// } |
282 | /// } |
283 | /// } |
284 | /// ``` |
285 | pub fn register(&self, waker: &Waker) { |
286 | match self |
287 | .state |
288 | .compare_exchange(WAITING, REGISTERING, Acquire, Acquire) |
289 | .unwrap_or_else(|x| x) |
290 | { |
291 | WAITING => { |
292 | unsafe { |
293 | // Locked acquired, update the waker cell |
294 | |
295 | // Avoid cloning the waker if the old waker will awaken the same task. |
296 | match &*self.waker.get() { |
297 | Some(old_waker) if old_waker.will_wake(waker) => (), |
298 | _ => *self.waker.get() = Some(waker.clone()), |
299 | } |
300 | |
301 | // Release the lock. If the state transitioned to include |
302 | // the `WAKING` bit, this means that at least one wake has |
303 | // been called concurrently. |
304 | // |
305 | // Start by assuming that the state is `REGISTERING` as this |
306 | // is what we just set it to. If this holds, we know that no |
307 | // other writes were performed in the meantime, so there is |
308 | // nothing to acquire, only release. In case of concurrent |
309 | // wakers, we need to acquire their releases, so success needs |
310 | // to do both. |
311 | let res = self |
312 | .state |
313 | .compare_exchange(REGISTERING, WAITING, AcqRel, Acquire); |
314 | |
315 | match res { |
316 | Ok(_) => { |
317 | // memory ordering: acquired self.state during CAS |
318 | // - if previous wakes went through it syncs with |
319 | // their final release (`fetch_and`) |
320 | // - if there was no previous wake the next wake |
321 | // will wake us, no sync needed. |
322 | } |
323 | Err(actual) => { |
324 | // This branch can only be reached if at least one |
325 | // concurrent thread called `wake`. In this |
326 | // case, `actual` **must** be `REGISTERING | |
327 | // `WAKING`. |
328 | debug_assert_eq!(actual, REGISTERING | WAKING); |
329 | |
330 | // Take the waker to wake once the atomic operation has |
331 | // completed. |
332 | let waker = (*self.waker.get()).take().unwrap(); |
333 | |
334 | // We need to return to WAITING state (clear our lock and |
335 | // concurrent WAKING flag). This needs to acquire all |
336 | // WAKING fetch_or releases and it needs to release our |
337 | // update to self.waker, so we need a `swap` operation. |
338 | self.state.swap(WAITING, AcqRel); |
339 | |
340 | // memory ordering: we acquired the state for all |
341 | // concurrent wakes, but future wakes might still |
342 | // need to wake us in case we can't make progress |
343 | // from the pending wakes. |
344 | // |
345 | // So we simply schedule to come back later (we could |
346 | // also simply leave the registration in place above). |
347 | waker.wake(); |
348 | } |
349 | } |
350 | } |
351 | } |
352 | WAKING => { |
353 | // Currently in the process of waking the task, i.e., |
354 | // `wake` is currently being called on the old task handle. |
355 | // |
356 | // memory ordering: we acquired the state for all |
357 | // concurrent wakes, but future wakes might still |
358 | // need to wake us in case we can't make progress |
359 | // from the pending wakes. |
360 | // |
361 | // So we simply schedule to come back later (we |
362 | // could also spin here trying to acquire the lock |
363 | // to register). |
364 | waker.wake_by_ref(); |
365 | } |
366 | state => { |
367 | // In this case, a concurrent thread is holding the |
368 | // "registering" lock. This probably indicates a bug in the |
369 | // caller's code as racing to call `register` doesn't make much |
370 | // sense. |
371 | // |
372 | // memory ordering: don't care. a concurrent register() is going |
373 | // to succeed and provide proper memory ordering. |
374 | // |
375 | // We just want to maintain memory safety. It is ok to drop the |
376 | // call to `register`. |
377 | debug_assert!(state == REGISTERING || state == REGISTERING | WAKING); |
378 | } |
379 | } |
380 | } |
381 | |
382 | /// Calls `wake` on the last `Waker` passed to `register`. |
383 | /// |
384 | /// If `register` has not been called yet, then this does nothing. |
385 | pub fn wake(&self) { |
386 | if let Some(waker) = self.take() { |
387 | waker.wake(); |
388 | } |
389 | } |
390 | |
391 | /// Returns the last `Waker` passed to `register`, so that the user can wake it. |
392 | /// |
393 | /// |
394 | /// Sometimes, just waking the AtomicWaker is not fine grained enough. This allows the user |
395 | /// to take the waker and then wake it separately, rather than performing both steps in one |
396 | /// atomic action. |
397 | /// |
398 | /// If a waker has not been registered, this returns `None`. |
399 | pub fn take(&self) -> Option<Waker> { |
400 | // AcqRel ordering is used in order to acquire the value of the `task` |
401 | // cell as well as to establish a `release` ordering with whatever |
402 | // memory the `AtomicWaker` is associated with. |
403 | match self.state.fetch_or(WAKING, AcqRel) { |
404 | WAITING => { |
405 | // The waking lock has been acquired. |
406 | let waker = unsafe { (*self.waker.get()).take() }; |
407 | |
408 | // Release the lock |
409 | self.state.fetch_and(!WAKING, Release); |
410 | |
411 | waker |
412 | } |
413 | state => { |
414 | // There is a concurrent thread currently updating the |
415 | // associated task. |
416 | // |
417 | // Nothing more to do as the `WAKING` bit has been set. It |
418 | // doesn't matter if there are concurrent registering threads or |
419 | // not. |
420 | // |
421 | debug_assert!( |
422 | state == REGISTERING || state == REGISTERING | WAKING || state == WAKING |
423 | ); |
424 | None |
425 | } |
426 | } |
427 | } |
428 | } |
429 | |
430 | impl Default for AtomicWaker { |
431 | fn default() -> Self { |
432 | AtomicWaker::new() |
433 | } |
434 | } |
435 | |
436 | impl fmt::Debug for AtomicWaker { |
437 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
438 | write!(f, "AtomicWaker" ) |
439 | } |
440 | } |
441 | |
442 | unsafe impl Send for AtomicWaker {} |
443 | unsafe impl Sync for AtomicWaker {} |
444 | |