Summary

Unix domain sockets provide a commonly used form of IPC on Unix-derived systems. This RFC proposes move the unix_socket nursery crate into the std::os::unix module.

Motivation

Unix sockets are a common form of IPC on unixy systems. Databases like PostgreSQL and Redis allow connections via Unix sockets, and Servo uses them to communicate with subprocesses. Even though Unix sockets are not present on Windows, their use is sufficiently widespread to warrant inclusion in the platform-specific sections of the standard library.

Detailed design

Unix sockets can be configured with the SOCK_STREAM, SOCK_DGRAM, and SOCK_SEQPACKET types. SOCK_STREAM creates a connection-oriented socket that behaves like a TCP socket, SOCK_DGRAM creates a packet-oriented socket that behaves like a UDP socket, and SOCK_SEQPACKET provides something of a hybrid between the other two - a connection-oriented, reliable, ordered stream of delimited packets. SOCK_SEQPACKET support has not yet been implemented in the unix_socket crate, so only the first two socket types will initially be supported in the standard library.

While a TCP or UDP socket would be identified by a IP address and port number, Unix sockets are typically identified by a filesystem path. For example, a Postgres server will listen on a Unix socket located at /run/postgresql/.s.PGSQL.5432 in some configurations. However, the socketpair function can make a pair of unnamed connected Unix sockets not associated with a filesystem path. In addition, Linux provides a separate abstract namespace not associated with the filesystem, indicated by a leading null byte in the address. In the initial implementation, the abstract namespace will not be supported - the various socket constructors will check for and reject addresses with interior null bytes.

A std::os::unix::net module will be created with the following contents:

The UnixStream type mirrors TcpStream:

pub struct UnixStream {
    ...
}

impl UnixStream {
    /// Connects to the socket named by `path`.
    ///
    /// `path` may not contain any null bytes.
    pub fn connect<P: AsRef<Path>>(path: P) -> io::Result<UnixStream> {
        ...
    }

    /// Creates an unnamed pair of connected sockets.
    ///
    /// Returns two `UnixStream`s which are connected to each other.
    pub fn pair() -> io::Result<(UnixStream, UnixStream)> {
        ...
    }

    /// Creates a new independently owned handle to the underlying socket.
    ///
    /// The returned `UnixStream` is a reference to the same stream that this
    /// object references. Both handles will read and write the same stream of
    /// data, and options set on one stream will be propagated to the other
    /// stream.
    pub fn try_clone(&self) -> io::Result<UnixStream> {
        ...
    }

    /// Returns the socket address of the local half of this connection.
    pub fn local_addr(&self) -> io::Result<SocketAddr> {
        ...
    }

    /// Returns the socket address of the remote half of this connection.
    pub fn peer_addr(&self) -> io::Result<SocketAddr> {
        ...
    }

    /// Sets the read timeout for the socket.
    ///
    /// If the provided value is `None`, then `read` calls will block
    /// indefinitely. It is an error to pass the zero `Duration` to this
    /// method.
    pub fn set_read_timeout(&self, timeout: Option<Duration>) -> io::Result<()> {
        ...
    }

    /// Sets the write timeout for the socket.
    ///
    /// If the provided value is `None`, then `write` calls will block
    /// indefinitely. It is an error to pass the zero `Duration` to this
    /// method.
    pub fn set_write_timeout(&self, timeout: Option<Duration>) -> io::Result<()> {
        ...
    }

    /// Returns the read timeout of this socket.
    pub fn read_timeout(&self) -> io::Result<Option<Duration>> {
        ...
    }

    /// Returns the write timeout of this socket.
    pub fn write_timeout(&self) -> io::Result<Option<Duration>> {
        ...
    }

    /// Moves the socket into or out of nonblocking mode.
    pub fn set_nonblocking(&self, nonblocking: bool) -> io::Result<()> {
        ...
    }

    /// Returns the value of the `SO_ERROR` option.
    pub fn take_error(&self) -> io::Result<Option<io::Error>> {
        ...
    }

    /// Shuts down the read, write, or both halves of this connection.
    ///
    /// This function will cause all pending and future I/O calls on the
    /// specified portions to immediately return with an appropriate value
    /// (see the documentation of `Shutdown`).
    pub fn shutdown(&self, how: Shutdown) -> io::Result<()> {
        ...
    }
}

impl Read for UnixStream {
    ...
}

impl<'a> Read for &'a UnixStream {
    ...
}

impl Write for UnixStream {
    ...
}

impl<'a> Write for UnixStream {
    ...
}

impl FromRawFd for UnixStream {
    ...
}

impl AsRawFd for UnixStream {
    ...
}

impl IntoRawFd for UnixStream {
    ...
}

Differences from TcpStream:

  • connect takes an AsRef<Path> rather than a ToSocketAddrs.
  • The pair method creates a pair of connected, unnamed sockets, as this is commonly used for IPC.
  • The SocketAddr returned by the local_addr and peer_addr methods is different.
  • The set_nonblocking and take_error methods are not currently present on TcpStream but are provided in the net2 crate and are being proposed for addition to the standard library in a separate RFC.

As noted above, a Unix socket can either be unnamed, be associated with a path on the filesystem, or (on Linux) be associated with an ID in the abstract namespace. The SocketAddr struct is fairly simple:

pub struct SocketAddr {
    ...
}

impl SocketAddr {
    /// Returns true if the address is unnamed.
    pub fn is_unnamed(&self) -> bool {
        ...
    }

    /// Returns the contents of this address if it corresponds to a filesystem path.
    pub fn as_pathname(&self) -> Option<&Path> {
        ...
    }
}

The UnixListener type mirrors the TcpListener type:

pub struct UnixListener {
    ...
}

impl UnixListener {
    /// Creates a new `UnixListener` bound to the specified socket.
    ///
    /// `path` may not contain any null bytes.
    pub fn bind<P: AsRef<Path>>(path: P) -> io::Result<UnixListener> {
        ...
    }

    /// Accepts a new incoming connection to this listener.
    ///
    /// This function will block the calling thread until a new Unix connection
    /// is established. When established, the corersponding `UnixStream` and
    /// the remote peer's address will be returned.
    pub fn accept(&self) -> io::Result<(UnixStream, SocketAddr)> {
        ...
    }

    /// Creates a new independently owned handle to the underlying socket.
    ///
    /// The returned `UnixListener` is a reference to the same socket that this
    /// object references. Both handles can be used to accept incoming
    /// connections and options set on one listener will affect the other.
    pub fn try_clone(&self) -> io::Result<UnixListener> {
        ...
    }

    /// Returns the local socket address of this listener.
    pub fn local_addr(&self) -> io::Result<SocketAddr> {
        ...
    }

    /// Moves the socket into or out of nonblocking mode.
    pub fn set_nonblocking(&self, nonblocking: bool) -> io::Result<()> {
        ...
    }

    /// Returns the value of the `SO_ERROR` option.
    pub fn take_error(&self) -> io::Result<Option<io::Error>> {
        ...
    }

    /// Returns an iterator over incoming connections.
    ///
    /// The iterator will never return `None` and will also not yield the
    /// peer's `SocketAddr` structure.
    pub fn incoming<'a>(&'a self) -> Incoming<'a> {
        ...
    }
}

impl FromRawFd for UnixListener {
    ...
}

impl AsRawFd for UnixListener {
    ...
}

impl IntoRawFd for UnixListener {
    ...
}

Differences from TcpListener:

  • bind takes an AsRef<Path> rather than a ToSocketAddrs.
  • The SocketAddr type is different.
  • The set_nonblocking and take_error methods are not currently present on TcpListener but are provided in the net2 crate and are being proposed for addition to the standard library in a separate RFC.

Finally, the UnixDatagram type mirrors the UpdSocket type:

pub struct UnixDatagram {
    ...
}

impl UnixDatagram {
    /// Creates a Unix datagram socket bound to the given path.
    ///
    /// `path` may not contain any null bytes.
    pub fn bind<P: AsRef<Path>>(path: P) -> io::Result<UnixDatagram> {
        ...
    }

    /// Creates a Unix Datagram socket which is not bound to any address.
    pub fn unbound() -> io::Result<UnixDatagram> {
        ...
    }

    /// Create an unnamed pair of connected sockets.
    ///
    /// Returns two `UnixDatagrams`s which are connected to each other.
    pub fn pair() -> io::Result<(UnixDatagram, UnixDatagram)> {
        ...
    }

    /// Creates a new independently owned handle to the underlying socket.
    ///
    /// The returned `UnixDatagram` is a reference to the same stream that this
    /// object references. Both handles will read and write the same stream of
    /// data, and options set on one stream will be propagated to the other
    /// stream.
    pub fn try_clone(&self) -> io::Result<UnixStream> {
        ...
    }

    /// Connects the socket to the specified address.
    ///
    /// The `send` method may be used to send data to the specified address.
    /// `recv` and `recv_from` will only receive data from that address.
    ///
    /// `path` may not contain any null bytes.
    pub fn connect<P: AsRef<Path>>(&self, path: P) -> io::Result<()> {
        ...
    }

    /// Returns the address of this socket.
    pub fn local_addr(&self) -> io::Result<SocketAddr> {
        ...
    }

    /// Returns the address of this socket's peer.
    ///
    /// The `connect` method will connect the socket to a peer.
    pub fn peer_addr(&self) -> io::Result<SocketAddr> {
        ...
    }

    /// Receives data from the socket.
    ///
    /// On success, returns the number of bytes read and the address from
    /// whence the data came.
    pub fn recv_from(&self, buf: &mut [u8]) -> io::Result<(usize, SocketAddr)> {
        ...
    }

    /// Receives data from the socket.
    ///
    /// On success, returns the number of bytes read.
    pub fn recv(&self, buf: &mut [u8]) -> io::Result<usize> {
        ...
    }

    /// Sends data on the socket to the specified address.
    ///
    /// On success, returns the number of bytes written.
    ///
    /// `path` may not contain any null bytes.
    pub fn send_to<P: AsRef<Path>>(&self, buf: &[u8], path: P) -> io::Result<usize> {
        ...
    }

    /// Sends data on the socket to the socket's peer.
    ///
    /// The peer address may be set by the `connect` method, and this method
    /// will return an error if the socket has not already been connected.
    ///
    /// On success, returns the number of bytes written.
    pub fn send(&self, buf: &[u8]) -> io::Result<usize> {
        ...
    }

    /// Sets the read timeout for the socket.
    ///
    /// If the provided value is `None`, then `recv` and `recv_from` calls will
    /// block indefinitely. It is an error to pass the zero `Duration` to this
    /// method.
    pub fn set_read_timeout(&self, timeout: Option<Duration>) -> io::Result<()> {
        ...
    }

    /// Sets the write timeout for the socket.
    ///
    /// If the provided value is `None`, then `send` and `send_to` calls will
    /// block indefinitely. It is an error to pass the zero `Duration` to this
    /// method.
    pub fn set_write_timeout(&self, timeout: Option<Duration>) -> io::Result<()> {
        ...
    }

    /// Returns the read timeout of this socket.
    pub fn read_timeout(&self) -> io::Result<Option<Duration>> {
        ...
    }

    /// Returns the write timeout of this socket.
    pub fn write_timeout(&self) -> io::Result<Option<Duration>> {
        ...
    }

    /// Moves the socket into or out of nonblocking mode.
    pub fn set_nonblocking(&self, nonblocking: bool) -> io::Result<()> {
        ...
    }

    /// Returns the value of the `SO_ERROR` option.
    pub fn take_error(&self) -> io::Result<Option<io::Error>> {
        ...
    }

    /// Shut down the read, write, or both halves of this connection.
    ///
    /// This function will cause all pending and future I/O calls on the
    /// specified portions to immediately return with an appropriate value
    /// (see the documentation of `Shutdown`).
    pub fn shutdown(&self, how: Shutdown) -> io::Result<()> {
        ...
    }
}

impl FromRawFd for UnixDatagram {
    ...
}

impl AsRawFd for UnixDatagram {
    ...
}

impl IntoRawFd for UnixDatagram {
    ...
}

Differences from UdpSocket:

  • bind takes an AsRef<Path> rather than a ToSocketAddrs.
  • The unbound method creates an unbound socket, as a Unix socket does not need to be bound to send messages.
  • The pair method creates a pair of connected, unnamed sockets, as this is commonly used for IPC.
  • The SocketAddr returned by the local_addr and peer_addr methods is different.
  • The connect, send, recv, set_nonblocking, and take_error methods are not currently present on UdpSocket but are provided in the net2 crate and are being proposed for addition to the standard library in a separate RFC.

Functionality not present

Some functionality is notably absent from this proposal:

  • Linux’s abstract namespace is not supported. Functionality may be added in the future via extension traits in std::os::linux::net.
  • No support for SOCK_SEQPACKET sockets is proposed, as it has not yet been implemented. Since it is connection oriented, there will be a socket type UnixSeqPacket and a listener type UnixSeqListener. The naming of the listener is a bit unfortunate, but use of SOCK_SEQPACKET is rare compared to SOCK_STREAM so naming priority can go to that version.
  • Unix sockets support file descriptor and credential transfer, but these will not initially be supported as the sendmsg/recvmsg interface is complex and bindings will need some time to prototype.

These features can bake in the rust-lang-nursery/unix-socket as they’re developed.

Drawbacks

While there is precedent for platform specific components in the standard library, this will be the by far the largest platform specific addition.

Alternatives

Unix socket support could be left out of tree.

The naming convention of UnixStream and UnixDatagram doesn’t perfectly mirror TcpStream and UdpSocket, but UnixStream and UnixSocket seems way too confusing.

Unresolved questions

Is std::os::unix::net the right name for this module? It’s not strictly “networking” as all communication is local to one machine. std::os::unix::unix is more accurate but weirdly repetitive and the extension trait module std::os::linux::unix is even weirder. std::os::unix::socket is an option, but seems like too general of a name for specifically AF_UNIX sockets as opposed to all sockets.