Struct ListenOptions

Source

pub struct ListenOptions { /* private fields */ }

Expand description

Options set on a listening socket before bind and listen.

use net_io::{ListenOptions, TcpListener};
let options = ListenOptions::new().backlog(4096).reuse_port(cfg!(target_os = "linux"));
let listener = TcpListener::bind_with("127.0.0.1:0", options)?;
println!("effective backlog {:?}", listener.info().backlog);

Implementations§

Source §

impl ListenOptions

Source

pub fn new() -> ListenOptions

Backlog 1,024; reuse address on Unix, exclusive address use on Windows; IPv6-only; TCP_NODELAY on accepted sockets.

Source

pub fn backlog(self, n: u32) -> ListenOptions

The listen backlog (capped by the OS; the effective value is in ListenerInfo).

Source

pub fn reuse_address(self, on: bool) -> ListenOptions

Unix: SO_REUSEADDR. Windows: false = SO_EXCLUSIVEADDRUSE, true = SO_REUSEADDR.

Source

pub fn reuse_port(self, on: bool) -> ListenOptions

SO_REUSEPORT (Unsupported on Windows).

Source

pub fn ipv6_only(self, on: bool) -> ListenOptions

IPV6_V6ONLY for IPv6 addresses (on by default; decision D-026).

Source

pub fn send_buffer(self, bytes: u32) -> ListenOptions

SO_SNDBUF on the listener (before listen).

Source

pub fn receive_buffer(self, bytes: u32) -> ListenOptions

SO_RCVBUF on the listener (before listen, so window scaling is negotiated).

Source

pub fn defer_accept(self, wait: Option<Duration>) -> ListenOptions

accept returns once the client has sent bytes, or after wait (TCP_DEFER_ACCEPT, whole seconds; Unsupported off Linux). A client that connects and stays silent costs no accept for that long. Cost: the gate sees the connection later, and a server that speaks first must not use this.

Source

pub fn fast_open(self, queue: Option<u32>) -> ListenOptions

Server TCP Fast Open with this many pending requests. Saves a round trip for a client that returns. Security: the first flight of data can be replayed by the network, so turn this on only for requests that are safe to run twice; the gate and the PROXY check run as usual. Linux also needs net.ipv4.tcp_fastopen.

Source

pub fn incoming_cpu(self, cpu: Option<u32>) -> ListenOptions

Accepts on this listener the connections the given CPU received (SO_INCOMING_CPU, Linux; Unsupported elsewhere). For one reuse-port listener per CPU.

Source

pub fn per_worker(self, on: bool) -> ListenOptions

One socket and one accept thread per reactor worker, in a SO_REUSEPORT group (Linux; Unsupported elsewhere, InvalidConfig on the blocking driver). Each feeds its own worker, so no connection crosses to another worker, and no accept thread is a bottleneck; with ServerBuilder::placement each socket also takes the connections its worker’s CPU received. Costs: a socket and a thread per worker, the kernel’s hash decides the balance, and such a listener cannot be handed off: closing one socket of a group resets what is queued on it.

Source