Implmement Happy Eyeballs RFC

src/eyeballs.rs

@@ -0,0 +1,163 @@
+//! A Happy Eyeballs RFC implementation
+//!
+//! Races interleaved IPv4 and IPv6 connections to provide the fastest connection
+//! in cases where certain addresses or address families might be blocked, broken, or slow.
+//! (See <https://datatracker.ietf.org/doc/html/rfc8305>)
+//!
+//! ureq strives for simplicity, and avoids spawning threads where it can,
+//! but - like with SOCKS - there's no way around it here.
+//! Some mini internal async executor
+//! (discussed in <https://github.com/algesten/ureq/issues/535#issuecomment-1229433311>)
+//! wouldn't help - `connect()` is a blocking syscall with no non-blocking alternative.
+//! (Big async runtimes like Tokio "solve" this problem by keeping a pool of OS threads
+//! around for just these sorts of blocking calls.)
+//! We _could_ have some thread pool (a la rayon) to avoid spawning threads
+//! on each connection attempt, but spawning a few threads is a cheap operation
+//! compared to everything else going on here.
+//! (DNS resolution, handshaking across the Internet...)
+//!
+//! Much of this implementation was inspired by attohttpc's:
+//! <https://github.com/sbstp/attohttpc/blob/master/src/happy.rs>
+
+use std::{
+    io,
+    iter::FusedIterator,
+    net::{SocketAddr, TcpStream},
+    sync::mpsc::{channel, RecvTimeoutError},
+    thread,
+    time::Instant,
+};
+
+use log::debug;
+
+use crate::timeout::{io_err_timeout, time_until_deadline};
+
+const TIMEOUT_MSG: &str = "timed out connecting";
+
+pub fn connect(
+    netloc: String,
+    addrs: &[SocketAddr],
+    deadline: Option<Instant>,
+) -> io::Result<(TcpStream, SocketAddr)> {
+    assert!(!addrs.is_empty());
+
+    // No racing needed if there's a single address.
+    if let [single] = addrs {
+        return single_connection(&netloc, *single, deadline);
+    }
+
+    // Interleave IPV4 and IPV6 addresses
+    let fours = addrs.iter().filter(|a| matches!(a, SocketAddr::V4(_)));
+    let sixes = addrs.iter().filter(|a| matches!(a, SocketAddr::V6(_)));
+    let sorted = interleave(fours, sixes);
+
+    let (tx, rx) = channel();
+    let mut first_error = None;
+
+    // Race connections!
+    // The RFC says:
+    //
+    // 1. Not to start connections "simultaneously", but since `connect()`
+    //    syscalls don't return until they've connected or timed out,
+    //    we don't have a way to start an attempt without blocking until it finishes.
+    //    (And if we did that, we wouldn't be racing!)
+    //
+    // 2. Once we have a successful connection, all other attempts should be cancelled.
+    //    Doing so would require a lot of nasty (and platform-specific) signal handling,
+    //    as it's the only way to interrupt `connect()`.
+    for s in sorted {
+        // Instead, make a best effort to not start new connections if we've got one already.
+        if let Ok(resp) = rx.try_recv() {
+            match resp {
+                Ok(c) => return Ok(c),
+                Err(e) => {
+                    let _ = first_error.get_or_insert(e);
+                }
+            }
+        }
+
+        let tx2 = tx.clone();
+        let nl2 = netloc.clone();
+        let s2 = *s;
+        thread::spawn(move || {
+            // If the receiver was dropped, someone else already won the race.
+            let _ = tx2.send(single_connection(&nl2, s2, deadline));
+        });
+    }
+    drop(tx);
+
+    const UNREACHABLE_MSG: &str =
+        "Unreachable: All Happy Eyeballs connections failed, but no error";
+
+    if let Some(d) = deadline {
+        // Wait for a successful connection, or for us to run out of time
+        loop {
+            let timeout = time_until_deadline(d, TIMEOUT_MSG)?;
+            match rx.recv_timeout(timeout) {
+                Ok(Ok(c)) => return Ok(c),
+                Ok(Err(e)) => {
+                    let _ = first_error.get_or_insert(e);
+                }
+                Err(RecvTimeoutError::Timeout) => {
+                    return Err(io_err_timeout(TIMEOUT_MSG.to_string()))
+                }
+                // If all the connecting threads hung up and none succeeded,
+                // return the first error.
+                Err(RecvTimeoutError::Disconnected) => {
+                    return Err(first_error.expect(UNREACHABLE_MSG))
+                }
+            };
+        }
+    } else {
+        // If there's no deadline, just wait around.
+        let connections = rx.iter();
+        for c in connections {
+            match c {
+                Ok(c) => return Ok(c),
+                Err(e) => {
+                    let _ = first_error.get_or_insert(e);
+                }
+            }
+        }
+        // If we got here, everyone failed. Return the first error.
+        Err(first_error.expect(UNREACHABLE_MSG))
+    }
+}
+
+fn single_connection(
+    netloc: &str,
+    addr: SocketAddr,
+    deadline: Option<Instant>,
+) -> io::Result<(TcpStream, SocketAddr)> {
+    debug!("connecting to {} at {}", netloc, addr);
+    if let Some(d) = deadline {
+        let timeout = time_until_deadline(d, TIMEOUT_MSG)?;
+        Ok((TcpStream::connect_timeout(&addr, timeout)?, addr))
+    } else {
+        Ok((TcpStream::connect(addr)?, addr))
+    }
+}
+
+fn interleave<T, A, B>(mut left: A, mut right: B) -> impl Iterator<Item = T>
+where
+    A: FusedIterator<Item = T>,
+    B: FusedIterator<Item = T>,
+{
+    let mut last_right = None;
+
+    std::iter::from_fn(move || {
+        if let Some(r) = last_right.take() {
+            return Some(r);
+        }
+
+        match (left.next(), right.next()) {
+            (Some(l), Some(r)) => {
+                last_right = Some(r);
+                Some(l)
+            }
+            (Some(l), None) => Some(l),
+            (None, Some(r)) => Some(r),
+            (None, None) => None,
+        }
+    })
+}

src/lib.rs

@@ -358,6 +358,7 @@ mod agent;
 mod body;
 mod chunked;
 mod error;
+mod eyeballs;
 mod header;
 mod middleware;
 mod pool;

src/response.rs

@@ -530,7 +530,7 @@ impl Response {
     /// ```
     #[cfg(feature = "json")]
     pub fn into_json<T: DeserializeOwned>(self) -> io::Result<T> {
-        use crate::stream::io_err_timeout;
+        use crate::timeout::io_err_timeout;
 
         let reader = self.into_reader();
         serde_json::from_reader(reader).map_err(|e| {

src/stream.rs

@@ -2,7 +2,6 @@ use log::debug;
 use std::io::{self, BufRead, BufReader, Read, Write};
 use std::net::SocketAddr;
 use std::net::TcpStream;
-use std::ops::Div;
 use std::time::Duration;
 use std::time::Instant;
 use std::{fmt, io::Cursor};
@@ -12,6 +11,7 @@ use socks::{TargetAddr, ToTargetAddr};
 
 use crate::chunked::Decoder as ChunkDecoder;
 use crate::error::ErrorKind;
+use crate::eyeballs;
 use crate::pool::{PoolKey, PoolReturner};
 use crate::proxy::Proxy;
 use crate::timeout::{io_err_timeout, time_until_deadline};
@@ -361,59 +361,35 @@ pub(crate) fn connect_host(
 
     let proto = proxy.as_ref().map(|proxy| proxy.proto);
 
-    let mut any_err = None;
-    let mut any_stream_and_addr = None;
-    // Find the first sock_addr that accepts a connection
-    let multiple_addrs = sock_addrs.len() > 1;
-
-    for sock_addr in sock_addrs {
-        // ensure connect timeout or overall timeout aren't yet hit.
-        let timeout = match connect_deadline {
-            Some(deadline) => {
-                let mut deadline = time_until_deadline(deadline, TIMEOUT_MSG)?;
-                if multiple_addrs {
-                    deadline = deadline.div(2);
-                }
-                Some(deadline)
-            }
-            None => None,
-        };
-
-        debug!("connecting to {} at {}", netloc, &sock_addr);
-
-        // connect with a configured timeout.
-        #[allow(clippy::unnecessary_unwrap)]
-        let stream = if proto.is_some() && Some(Proto::HTTP) != proto {
-            connect_socks(
+    let (mut stream, remote_addr) = if proto.is_some() && Some(Proto::HTTP) != proto {
+        // SOCKS proxy connections.
+        // Don't mix that with happy eyeballs
+        // (where we race multiple connections and take the fastest)
+        // since we'd be repeatedly connecting to the same proxy server.
+        let mut stream_and_addr_result = None;
+        // Find the first sock_addr that accepts a connection
+        for sock_addr in sock_addrs {
+            // ensure connect timeout or overall timeout aren't yet hit.
+            debug!("connecting to {} at {}", netloc, &sock_addr);
+
+            // connect with a configured timeout.
+            #[allow(clippy::unnecessary_unwrap)]
+            let stream = connect_socks(
                 unit,
                 proxy.clone().unwrap(),
                 connect_deadline,
                 sock_addr,
                 hostname,
                 port,
                 proto.unwrap(),
-            )
-        } else if let Some(timeout) = timeout {
-            TcpStream::connect_timeout(&sock_addr, timeout)
-        } else {
-            TcpStream::connect(sock_addr)
-        };
-
-        if let Ok(stream) = stream {
-            any_stream_and_addr = Some((stream, sock_addr));
-            break;
-        } else if let Err(err) = stream {
-            any_err = Some(err);
+            );
+            stream_and_addr_result = Some(stream.map(|s| (s, sock_addr)));
         }
-    }
-
-    let (mut stream, remote_addr) = if let Some(stream_and_addr) = any_stream_and_addr {
-        stream_and_addr
-    } else if let Some(e) = any_err {
-        return Err(ErrorKind::ConnectionFailed.msg("Connect error").src(e));
+        stream_and_addr_result.expect("unreachable: connected to IPs, but no result")
     } else {
-        panic!("shouldn't happen: failed to connect to all IPs, but no error");
-    };
+        eyeballs::connect(netloc, &sock_addrs, connect_deadline)
+    }
+    .map_err(|e| ErrorKind::ConnectionFailed.msg("Connect error").src(e))?;
 
     stream.set_nodelay(unit.agent.config.no_delay)?;