1 | #![cfg_attr (not(feature = "sync" ), allow(dead_code, unreachable_pub))] |
2 | |
3 | //! A multi-producer, single-consumer queue for sending values between |
4 | //! asynchronous tasks. |
5 | //! |
6 | //! This module provides two variants of the channel: bounded and unbounded. The |
7 | //! bounded variant has a limit on the number of messages that the channel can |
8 | //! store, and if this limit is reached, trying to send another message will |
9 | //! wait until a message is received from the channel. An unbounded channel has |
10 | //! an infinite capacity, so the `send` method will always complete immediately. |
11 | //! This makes the [`UnboundedSender`] usable from both synchronous and |
12 | //! asynchronous code. |
13 | //! |
14 | //! Similar to the `mpsc` channels provided by `std`, the channel constructor |
15 | //! functions provide separate send and receive handles, [`Sender`] and |
16 | //! [`Receiver`] for the bounded channel, [`UnboundedSender`] and |
17 | //! [`UnboundedReceiver`] for the unbounded channel. If there is no message to read, |
18 | //! the current task will be notified when a new value is sent. [`Sender`] and |
19 | //! [`UnboundedSender`] allow sending values into the channel. If the bounded |
20 | //! channel is at capacity, the send is rejected and the task will be notified |
21 | //! when additional capacity is available. In other words, the channel provides |
22 | //! backpressure. |
23 | //! |
24 | //! This channel is also suitable for the single-producer single-consumer |
25 | //! use-case. (Unless you only need to send one message, in which case you |
26 | //! should use the [oneshot] channel.) |
27 | //! |
28 | //! # Disconnection |
29 | //! |
30 | //! When all [`Sender`] handles have been dropped, it is no longer |
31 | //! possible to send values into the channel. This is considered the termination |
32 | //! event of the stream. As such, `Receiver::poll` returns `Ok(Ready(None))`. |
33 | //! |
34 | //! If the [`Receiver`] handle is dropped, then messages can no longer |
35 | //! be read out of the channel. In this case, all further attempts to send will |
36 | //! result in an error. Additionally, all unread messages will be drained from the |
37 | //! channel and dropped. |
38 | //! |
39 | //! # Clean Shutdown |
40 | //! |
41 | //! When the [`Receiver`] is dropped, it is possible for unprocessed messages to |
42 | //! remain in the channel. Instead, it is usually desirable to perform a "clean" |
43 | //! shutdown. To do this, the receiver first calls `close`, which will prevent |
44 | //! any further messages to be sent into the channel. Then, the receiver |
45 | //! consumes the channel to completion, at which point the receiver can be |
46 | //! dropped. |
47 | //! |
48 | //! # Communicating between sync and async code |
49 | //! |
50 | //! When you want to communicate between synchronous and asynchronous code, there |
51 | //! are two situations to consider: |
52 | //! |
53 | //! **Bounded channel**: If you need a bounded channel, you should use a bounded |
54 | //! Tokio `mpsc` channel for both directions of communication. Instead of calling |
55 | //! the async [`send`][bounded-send] or [`recv`][bounded-recv] methods, in |
56 | //! synchronous code you will need to use the [`blocking_send`][blocking-send] or |
57 | //! [`blocking_recv`][blocking-recv] methods. |
58 | //! |
59 | //! **Unbounded channel**: You should use the kind of channel that matches where |
60 | //! the receiver is. So for sending a message _from async to sync_, you should |
61 | //! use [the standard library unbounded channel][std-unbounded] or |
62 | //! [crossbeam][crossbeam-unbounded]. Similarly, for sending a message _from sync |
63 | //! to async_, you should use an unbounded Tokio `mpsc` channel. |
64 | //! |
65 | //! Please be aware that the above remarks were written with the `mpsc` channel |
66 | //! in mind, but they can also be generalized to other kinds of channels. In |
67 | //! general, any channel method that isn't marked async can be called anywhere, |
68 | //! including outside of the runtime. For example, sending a message on a |
69 | //! [oneshot] channel from outside the runtime is perfectly fine. |
70 | //! |
71 | //! # Multiple runtimes |
72 | //! |
73 | //! The `mpsc` channel is runtime agnostic. You can freely move it between |
74 | //! different instances of the Tokio runtime or even use it from non-Tokio |
75 | //! runtimes. |
76 | //! |
77 | //! When used in a Tokio runtime, it participates in |
78 | //! [cooperative scheduling](crate::task::coop#cooperative-scheduling) to avoid |
79 | //! starvation. This feature does not apply when used from non-Tokio runtimes. |
80 | //! |
81 | //! As an exception, methods ending in `_timeout` are not runtime agnostic |
82 | //! because they require access to the Tokio timer. See the documentation of |
83 | //! each `*_timeout` method for more information on its use. |
84 | //! |
85 | //! # Allocation behavior |
86 | //! |
87 | //! <div class="warning">The implementation details described in this section may change in future |
88 | //! Tokio releases.</div> |
89 | //! |
90 | //! The mpsc channel stores elements in blocks. Blocks are organized in a linked list. Sending |
91 | //! pushes new elements onto the block at the front of the list, and receiving pops them off the |
92 | //! one at the back. A block can hold 32 messages on a 64-bit target and 16 messages on a 32-bit |
93 | //! target. This number is independent of channel and message size. Each block also stores 4 |
94 | //! pointer-sized values for bookkeeping (so on a 64-bit machine, each message has 1 byte of |
95 | //! overhead). |
96 | //! |
97 | //! When all values in a block have been received, it becomes empty. It will then be freed, unless |
98 | //! the channel's first block (where newly-sent elements are being stored) has no next block. In |
99 | //! that case, the empty block is reused as the next block. |
100 | //! |
101 | //! [`Sender`]: crate::sync::mpsc::Sender |
102 | //! [`Receiver`]: crate::sync::mpsc::Receiver |
103 | //! [bounded-send]: crate::sync::mpsc::Sender::send() |
104 | //! [bounded-recv]: crate::sync::mpsc::Receiver::recv() |
105 | //! [blocking-send]: crate::sync::mpsc::Sender::blocking_send() |
106 | //! [blocking-recv]: crate::sync::mpsc::Receiver::blocking_recv() |
107 | //! [`UnboundedSender`]: crate::sync::mpsc::UnboundedSender |
108 | //! [`UnboundedReceiver`]: crate::sync::mpsc::UnboundedReceiver |
109 | //! [oneshot]: crate::sync::oneshot |
110 | //! [`Handle::block_on`]: crate::runtime::Handle::block_on() |
111 | //! [std-unbounded]: std::sync::mpsc::channel |
112 | //! [crossbeam-unbounded]: https://docs.rs/crossbeam/*/crossbeam/channel/fn.unbounded.html |
113 | //! [`send_timeout`]: crate::sync::mpsc::Sender::send_timeout |
114 | |
115 | pub(super) mod block; |
116 | |
117 | mod bounded; |
118 | pub use self::bounded::{ |
119 | channel, OwnedPermit, Permit, PermitIterator, Receiver, Sender, WeakSender, |
120 | }; |
121 | |
122 | mod chan; |
123 | |
124 | pub(super) mod list; |
125 | |
126 | mod unbounded; |
127 | pub use self::unbounded::{ |
128 | unbounded_channel, UnboundedReceiver, UnboundedSender, WeakUnboundedSender, |
129 | }; |
130 | |
131 | pub mod error; |
132 | |
133 | /// The number of values a block can contain. |
134 | /// |
135 | /// This value must be a power of 2. It also must be smaller than the number of |
136 | /// bits in `usize`. |
137 | #[cfg (all(target_pointer_width = "64" , not(loom)))] |
138 | const BLOCK_CAP: usize = 32; |
139 | |
140 | #[cfg (all(not(target_pointer_width = "64" ), not(loom)))] |
141 | const BLOCK_CAP: usize = 16; |
142 | |
143 | #[cfg (loom)] |
144 | const BLOCK_CAP: usize = 2; |
145 | |