|
| 1 | +- Feature Name: socket_timeouts |
| 2 | +- Start Date: 2015-04-08 |
| 3 | +- RFC PR: (leave this empty) |
| 4 | +- Rust Issue: (leave this empty) |
| 5 | + |
| 6 | +# Summary |
| 7 | + |
| 8 | +Add sockopt-style timeouts to `std::net` types. |
| 9 | + |
| 10 | +# Motivation |
| 11 | + |
| 12 | +Currently, operations on various socket types in `std::net` block |
| 13 | +indefinitely (i.e., until the connection is closed or data is |
| 14 | +transferred). But there are many contexts in which timing out a |
| 15 | +blocking call is important. |
| 16 | + |
| 17 | +The [goal of the current IO system][io-reform] is to gradually expose |
| 18 | +cross-platform, blocking APIs for IO, especially APIs that directly |
| 19 | +correspond to the underlying system APIs. Sockets are widely available |
| 20 | +with nearly identical system APIs across the platforms Rust targets, |
| 21 | +and this includes support for timeouts via [sockopts][sockopt]. |
| 22 | + |
| 23 | +So timeouts are well-motivated and well-suited to `std::net`. |
| 24 | + |
| 25 | +# Detailed design |
| 26 | + |
| 27 | +The proposal is to *directly expose* the timeout functionality |
| 28 | +provided by [`setsockopt`][sockopt], in much the same way we currently |
| 29 | +expose functionality like `set_nodelay`: |
| 30 | + |
| 31 | +```rust |
| 32 | +impl TcpStream { |
| 33 | + pub fn set_read_timeout(&self, dur: Duration) -> io::Result<()> { ... } |
| 34 | + pub fn set_write_timeout(&self, dur: Duration) -> io::Result<()> { ... } |
| 35 | +} |
| 36 | + |
| 37 | +impl UdpSocket { |
| 38 | + pub fn set_read_timeout(&self, dur: Duration) -> io::Result<()> { ... } |
| 39 | + pub fn set_write_timeout(&self, dur: Duration) -> io::Result<()> { ... } |
| 40 | +} |
| 41 | +``` |
| 42 | + |
| 43 | +These methods take an amount of time in the form of a `Duration`, |
| 44 | +which is [undergoing stabilization][duration-reform]. They are |
| 45 | +implemented via straightforward calls to `setsockopt`. |
| 46 | + |
| 47 | +# Drawbacks |
| 48 | + |
| 49 | +One potential downside to this design is that the timeouts are set |
| 50 | +through direct mutation of the socket state, which can lead to |
| 51 | +composition problems. For example, a socket could be passed to another |
| 52 | +function which needs to use it with a timeout, but setting the timeout |
| 53 | +clobbers any previous values. This lack of composability leads to |
| 54 | +defensive programming in the form of "callee save" resets of timeouts, |
| 55 | +for example. An alternative design is given below. |
| 56 | + |
| 57 | +The advantage of binding the mutating APIs directly is that we keep a |
| 58 | +close correspondence between the `std::net` types and their underlying |
| 59 | +system types, and a close correspondence between Rust APIs and system |
| 60 | +APIs. It's not clear that this kind of composability is important |
| 61 | +enough in practice to justify a departure from the traditional API. |
| 62 | + |
| 63 | +# Alternatives |
| 64 | + |
| 65 | +A different approach would be to *wrap* socket types with a "timeout |
| 66 | +modifier", which would be responsible for setting and resetting the |
| 67 | +timeouts: |
| 68 | + |
| 69 | +```rust |
| 70 | +struct WithTimeout<T> { |
| 71 | + timeout: Duration, |
| 72 | + innter: T |
| 73 | +} |
| 74 | + |
| 75 | +impl<T> WithTimeout<T> { |
| 76 | + /// Returns the wrapped object, resetting the timeout |
| 77 | + pub fn into_inner(self) -> T { ... } |
| 78 | +} |
| 79 | + |
| 80 | +impl TcpStream { |
| 81 | + /// Wraps the stream with a timeout |
| 82 | + pub fn with_timeout(self, timeout: Duration) -> WithTimeout<TcpStream> { ... } |
| 83 | +} |
| 84 | + |
| 85 | +impl<T: Read> Read for WithTimeout<T> { ... } |
| 86 | +impl<T: Write> Write for WithTimeout<T> { ... } |
| 87 | +``` |
| 88 | + |
| 89 | +A [previous RFC][deadlines] spelled this out in more detail. |
| 90 | + |
| 91 | +Unfortunately, such a "wrapping" API has problems of its own. It |
| 92 | +creates unfortunate type incompatibilities, since you cannot store a |
| 93 | +timeout-wrapped socket where a "normal" socket is expected. It is |
| 94 | +difficult to be "polymorphic" over timeouts. |
| 95 | + |
| 96 | +Ultimately, it's not clear that the extra complexities of the type |
| 97 | +distinction here are worth the better theoretical composability. |
| 98 | + |
| 99 | +# Unresolved questions |
| 100 | + |
| 101 | +Should we consider a preliminary version of this RFC that introduces |
| 102 | +methods like `set_read_timeout_ms`, similar to `wait_timeout_ms` on |
| 103 | +`Condvar`? These methods have been introduced elsewhere to provide a |
| 104 | +stable way to use timeouts prior to `Duration` being stabilized. |
| 105 | + |
| 106 | +[io-reform]: https://github.com/rust-lang/rfcs/blob/master/text/0517-io-os-reform.md |
| 107 | +[sockopt]: http://pubs.opengroup.org/onlinepubs/009695399/functions/setsockopt.html |
| 108 | +[duration-reform]: https://github.com/rust-lang/rfcs/pull/1040 |
| 109 | +[deadlines]: https://github.com/rust-lang/rfcs/pull/577/ |
0 commit comments