path: root/crates/tor-rtcompat/src/impls.rs

//! Different implementations of a common async API for use in arti
//!
//! Currently only async_std, tokio and smol are provided.

#[cfg(feature = "async-std")]
pub(crate) mod async_std;

#[cfg(feature = "tokio")]
pub(crate) mod tokio;

#[cfg(feature = "smol")]
pub(crate) mod smol;

#[cfg(feature = "rustls")]
pub(crate) mod rustls;

#[cfg(feature = "native-tls")]
pub(crate) mod native_tls;

pub(crate) mod streamops;
pub(crate) mod unimpl_tls;

use crate::network::{
    CommonConnectOptions, CommonListenOptions, TcpConnectOptions, TcpListenOptions,
};
use socket2::Socket;

#[cfg(unix)]
use tor_error::warn_report;

/// Connection backlog size to use for `listen()` calls on IP sockets.
//
// How this was chosen:
//
// 1. The rust standard library uses a backlog of 128 for TCP sockets. This matches `SOMAXCONN` on
//    most systems.
//
// 2. Mio (backend for tokio) previously used 1024. But they recently (confusingly) copied the logic
//    from the standard library's unix socket implementation, which uses different values on
//    different platforms. These values were tuned for unix sockets, so I think we should ignore
//    them and mio's implementation here.
//    https://github.com/tokio-rs/mio/pull/1896
//
// 3. Tor first tries using `INT_MAX`, and if that fails falls back to `SOMAXCONN` (using a global
//    to remember if it did the fallback for future listen() calls; see `tor_listen`).
//
// 4. On supported platforms, if you use a backlog that is too large, the system will supposedly
//    silently cap the value instead of failing.
//
//     Linux:
//     listen(2)
//     > If the backlog argument is greater than the value in /proc/sys/net/core/somaxconn, then it
//     > is silently capped to that value.
//
//     FreeBSD:
//     listen(2)
//     > The sysctl(3) MIB variable kern.ipc.soacceptqueue specifies a hard limit on backlog; if a
//     > value greater than kern.ipc.soacceptqueue or less than zero is specified, backlog is
//     > silently forced to kern.ipc.soacceptqueue.
//
//     OpenBSD:
//     listen(2)
//     > [BUGS] The backlog is currently limited (silently) to the value of the kern.somaxconn
//     > sysctl, which defaults to 128.
//
//     Windows:
//     https://learn.microsoft.com/en-us/windows/win32/api/winsock2/nf-winsock2-listen
//     > The backlog parameter is limited (silently) to a reasonable value as determined by the
//     > underlying service provider. Illegal values are replaced by the nearest legal value.
//
//     Mac OS:
//     Archived listen(2) docs
//     https://developer.apple.com/library/archive/documentation/System/Conceptual/ManPages_iPhoneOS/man2/listen.2.html
//     > [BUGS] The backlog is currently limited (silently) to 128.
//
// 5. While the rust APIs take a `u32`, the libc API uses `int`. So we shouldn't use a value larger
//    than `c_int::MAX`.
//
// 6. We should be careful not to set this too large, as supposedly some systems will truncate this to
//    16 bits. So for example a value of `65536` would cause a backlog of 1. But maybe they are just
//    referring to systems where `int` is 16 bits?
//    https://bugs.python.org/issue38699#msg357957
//
// Here we use `u16::MAX`. We assume that this will succeed on all supported platforms. Unlike tor,
// we do not try again with a smaller value since this doesn't seem to be needed on modern systems.
// We can add it if we find that it's needed.
//
// A value of `u16::MAX` is arguably too high, since a smaller value like 4096 would be large enough
// for legitimate traffic, and illegitimate traffic would be better handled by the kernel with
// something like SYN cookies. But it's easier for users to reduce the max using
// `/proc/sys/net/core/somaxconn` than to increase this max by recompiling arti.
const LISTEN_BACKLOG: i32 = u16::MAX as i32;

/// Open a listening TCP socket.
///
/// The socket will be non-blocking, and the socket handle will be close-on-exec/non-inheritable.
/// Other socket options may also be set depending on the socket type and platform.
///
/// Historically we relied on the runtime to create a listening socket, but we need some specific
/// socket options set, and not all runtimes will behave the same. It's better for us to create the
/// socket with the options we need and with consistent behaviour across all runtimes. For example
/// if each runtime were using a different `listen()` backlog size, it might be difficult to debug
/// related issues.
#[cfg(not(all(target_arch = "wasm32", target_os = "unknown")))]
pub(crate) fn tcp_listen(
    addr: &std::net::SocketAddr,
    options: &TcpListenOptions,
) -> std::io::Result<std::net::TcpListener> {
    use socket2::{Domain, Type};

    // Destructure the options so that we don't forget to use any.
    let TcpListenOptions {
        common:
            CommonListenOptions {
                send_buffer_size,
                recv_buffer_size,
            },
    } = options;

    // `socket2::Socket::new()`:
    // > This function corresponds to `socket(2)` on Unix and `WSASocketW` on Windows.
    // >
    // > On Unix-like systems, the close-on-exec flag is set on the new socket. Additionally, on
    // > Apple platforms `SOCK_NOSIGPIPE` is set. On Windows, the socket is made non-inheritable.
    let socket = match addr {
        std::net::SocketAddr::V4(_) => Socket::new(Domain::IPV4, Type::STREAM, None)?,
        std::net::SocketAddr::V6(_) => {
            let socket = Socket::new(Domain::IPV6, Type::STREAM, None)?;

            // On `cfg(unix)` systems, set `IPV6_V6ONLY` so that we can bind AF_INET and
            // AF_INET6 sockets to the same port.
            // This is `cfg(unix)` as I'm not sure what the socket option does (if anything) on
            // non-unix platforms.
            #[cfg(unix)]
            if let Err(e) = socket.set_only_v6(true) {
                // If we see this, we should exclude more platforms.
                warn_report!(
                    e,
                    "Failed to set `IPV6_V6ONLY` on `AF_INET6` socket. \
                    Please report this bug at https://gitlab.torproject.org/tpo/core/arti/-/issues",
                );
            }

            socket
        }
    };

    // Below we try to match what a `tokio::net::TcpListener::bind()` would do. This is a bit tricky
    // since tokio documents "Calling TcpListener::bind("127.0.0.1:8080") is equivalent to:" with
    // some provided example code, but this logic actually appears to happen in the mio crate, and
    // doesn't match exactly with tokio's documentation. So here we acknowledge that we likely do
    // deviate from `tokio::net::TcpListener::bind()` a bit.

    socket.set_nonblocking(true)?;

    // The docs for `tokio::net::TcpSocket` say:
    //
    // > // On platforms with Berkeley-derived sockets, this allows to quickly
    // > // rebind a socket, without needing to wait for the OS to clean up the
    // > // previous one.
    // >
    // > // On Windows, this allows rebinding sockets which are actively in use,
    // > // which allows "socket hijacking", so we explicitly don't set it here.
    // > // https://docs.microsoft.com/en-us/windows/win32/winsock/using-so-reuseaddr-and-so-exclusiveaddruse
    //
    // This appears to be a comment that tokio copied from mio.
    //
    // So here we only set SO_REUSEADDR for `cfg(unix)` to match tokio.
    #[cfg(unix)]
    socket.set_reuse_address(true)?;

    // tcp(7):
    //
    // > On individual connections, the socket buffer size must be set prior to the listen(2) or
    // > connect(2) calls in order to have it take effect.
    if let Some(send_buffer_size) = send_buffer_size {
        socket.set_send_buffer_size(*send_buffer_size)?;
    }
    if let Some(recv_buffer_size) = recv_buffer_size {
        socket.set_recv_buffer_size(*recv_buffer_size)?;
    }

    socket.bind(&(*addr).into())?;

    socket.listen(LISTEN_BACKLOG)?;

    Ok(socket.into())
}

/// Stub replacement for tcp_listen on wasm32-unknown
#[cfg(all(target_arch = "wasm32", target_os = "unknown"))]
pub(crate) fn tcp_listen(
    _addr: &std::net::SocketAddr,
    _options: &TcpListenOptions,
) -> std::io::Result<std::net::TcpListener> {
    Err(std::io::Error::from(std::io::ErrorKind::Unsupported))
}

/// Initialize a TCP socket in preparation for a connect().
///
/// The socket will be non-blocking, and the socket handle will be close-on-exec/non-inheritable.
/// Other socket options may also be set depending on the socket type and platform.
///
/// This returns a socket without any `connect()` call. The caller MUST:
///
/// 1. connect() the socket.
/// 2. Wait for the socket to become writable using whatever mechanism
///    is available with the current runtime.
/// 3. Check `SO_ERROR` for errors.
///
/// Historically we relied on the runtime to create and connect the socket, but we need some
/// specific socket options set, and not all runtimes will behave the same. It's better for us to
/// create the socket with the options we need and with consistent behaviour across all runtimes.
#[cfg(not(all(target_arch = "wasm32", target_os = "unknown")))]
pub(crate) fn tcp_pre_connect(
    addr: &std::net::SocketAddr,
    options: &TcpConnectOptions,
) -> std::io::Result<socket2::Socket> {
    use socket2::{Domain, Type};

    // Destructure the options so that we don't forget to use any.
    let TcpConnectOptions {
        common:
            CommonConnectOptions {
                send_buffer_size,
                recv_buffer_size,
            },
    } = options;

    let domain = match addr {
        std::net::SocketAddr::V4(_) => Domain::IPV4,
        std::net::SocketAddr::V6(_) => Domain::IPV6,
    };

    // `socket2::Socket::new()`:
    // > This function corresponds to `socket(2)` on Unix and `WSASocketW` on Windows.
    // >
    // > On Unix-like systems, the close-on-exec flag is set on the new socket. Additionally, on
    // > Apple platforms `SOCK_NOSIGPIPE` is set. On Windows, the socket is made non-inheritable.
    let socket = Socket::new(domain, Type::STREAM, None)?;

    socket.set_nonblocking(true)?;

    // tcp(7):
    //
    // > On individual connections, the socket buffer size must be set prior to the listen(2) or
    // > connect(2) calls in order to have it take effect.
    if let Some(send_buffer_size) = send_buffer_size {
        socket.set_send_buffer_size(*send_buffer_size)?;
    }
    if let Some(recv_buffer_size) = recv_buffer_size {
        socket.set_recv_buffer_size(*recv_buffer_size)?;
    }

    // TODO: In the future, we'll likely want to support optionally binding to an address or to a
    // network interface (`SO_BINDTODEVICE`). See c-tor's `OutboundBindAddresses`.
    // If we do, we will also want to set `IP_BIND_ADDRESS_NO_PORT`.
    // We may also want to consider setting `IPV6_V6ONLY` (do we want to support connecting to
    // IPv4-mapped IPv6 addresses while we already do happy eyeballs?).

    // We do not connect() here so that we can use whatever connection mechanism is best for the
    // runtime being used.

    Ok(socket)
}

/// Stub replacement for tcp_pre_connect on wasm32-unknown
#[cfg(all(target_arch = "wasm32", target_os = "unknown"))]
pub(crate) fn tcp_pre_connect(
    _addr: &std::net::SocketAddr,
    _options: &TcpConnectOptions,
) -> std::io::Result<socket2::Socket> {
    Err(std::io::Error::from(std::io::ErrorKind::Unsupported))
}

/// Connect a TCP socket using the async-io crate.
///
/// This in theory should be runtime-independent as async-io spawns its own thread to poll the
/// socket. But this is inefficient on some runtimes like tokio.
///
/// Runtimes that want to connect manually should use [`tcp_pre_connect()`] to set up the socket,
/// and then connect it manually.
#[cfg(any(feature = "async-std", feature = "smol"))]
#[cfg(not(all(target_arch = "wasm32", target_os = "unknown")))]
pub(crate) async fn tcp_async_io_connect(
    addr: &std::net::SocketAddr,
    options: &TcpConnectOptions,
) -> std::io::Result<std::net::TcpStream> {
    use async_io::Async;

    // The socket before connect() has been called.
    let socket = tcp_pre_connect(addr, options)?;

    // Different platforms return different results from non-blocking `connect()`s.
    // Here we've checked that we match mio (tokio's low-level I/O code) for unix and windows
    // to ensure that we're handling the right error kind/errno.
    match socket.connect(&(*addr).into()) {
        Ok(()) => {}
        // On unix, mio checks for `EINPROGRESS`:
        // https://github.com/tokio-rs/mio/blob/0db25a7eae653f02e964a28d9aaf65b74c941208/src/sys/unix/tcp.rs#L35
        #[cfg(unix)]
        Err(e) if e.raw_os_error() == Some(libc::EINPROGRESS) => {}
        // On windows, mio checks for `WouldBlock`:
        // https://github.com/tokio-rs/mio/blob/0db25a7eae653f02e964a28d9aaf65b74c941208/src/sys/windows/tcp.rs#L44
        #[cfg(windows)]
        Err(e) if e.kind() == std::io::ErrorKind::WouldBlock => {}
        Err(e) => return Err(e),
    }

    // The socket is already non-blocking,
    // so `Async` doesn't need to set as non-blocking again.
    let socket = Async::new_nonblocking(socket)?;

    // Wait for the socket to become writable, indicating that it's connected.
    socket.writable().await?;

    // Check `SO_ERROR`.
    if let Some(e) = socket.get_ref().take_error()? {
        return Err(e);
    }

    Ok(socket.into_inner()?.into())
}

/// Stub replacement for tcp_async_io_connect on wasm32-unknown
#[cfg(any(feature = "async-std", feature = "smol"))]
#[cfg(all(target_arch = "wasm32", target_os = "unknown"))]
pub(crate) async fn tcp_async_io_connect(
    _addr: &std::net::SocketAddr,
    _options: &TcpConnectOptions,
) -> std::io::Result<std::net::TcpStream> {
    Err(std::io::Error::from(std::io::ErrorKind::Unsupported))
}

/// Helper: Implement an unreachable NetProvider<unix::SocketAddr> for a given runtime.
#[cfg(not(unix))]
macro_rules! impl_unix_non_provider {
    { $for_type:ty } => {

        #[async_trait]
        impl crate::traits::NetStreamProvider<tor_general_addr::unix::SocketAddr> for $for_type {
            type Stream = crate::unimpl::FakeStream;
            type Listener = crate::unimpl::FakeListener<tor_general_addr::unix::SocketAddr>;
            type ConnectOptions = crate::network::UnixConnectOptions;
            type ListenOptions = crate::network::UnixListenOptions;
            async fn connect(
                &self,
                _a: &tor_general_addr::unix::SocketAddr,
                _options: &Self::ConnectOptions,
            ) -> IoResult<Self::Stream> {
                Err(tor_general_addr::unix::NoAfUnixSocketSupport::default().into())

            }
            async fn listen(
                &self,
                _a: &tor_general_addr::unix::SocketAddr,
                _options: &Self::ListenOptions,
            ) -> IoResult<Self::Listener> {
                Err(tor_general_addr::unix::NoAfUnixSocketSupport::default().into())
            }
        }
    }
}
#[cfg(not(unix))]
pub(crate) use impl_unix_non_provider;