//! Linux networking backend.

use std::cell::Cell;
use std::ffi::c_void;
use std::future::Future;
use std::io;
use std::mem::MaybeUninit;
use std::net::{
    Ipv4Addr, Ipv6Addr, Shutdown, SocketAddr, SocketAddrV4, SocketAddrV6, ToSocketAddrs,
};
use std::os::fd::{AsRawFd, FromRawFd, OwnedFd, RawFd};
use std::pin::Pin;
use std::thread;
use std::time::Duration;

thread_local! {
    // None = untested, Some(true) = io_uring works, Some(false) = use offload.
    // After the first successful IORING_OP_SEND, the clone in send() is skipped
    // for all subsequent calls on the same thread.
    static SEND_URING_SUPPORTED: Cell<Option<bool>> = const { Cell::new(None) };
}

use crate::op::completion::completion_for_current_thread;
use crate::op::net::{AcceptedSocket, NetOp, ReceivedDatagram};
use crate::platform::linux_x86_64::runtime::with_current_driver;
use crate::platform::linux_x86_64::uring::{
    IORING_OP_ACCEPT, IORING_OP_BIND, IORING_OP_CLOSE, IORING_OP_CONNECT, IORING_OP_LISTEN,
    IORING_OP_RECV, IORING_OP_RECVMSG, IORING_OP_SEND, IORING_OP_SENDMSG, IORING_OP_SHUTDOWN,
    IORING_OP_SOCKET, IoUringCqe, IoUringSqe,
};

const DEFAULT_LISTENER_BACKLOG: i32 = 1024;

#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum ExecutionPath {
    IoUring,
    Offload,
}

pub fn execution_path(op: &NetOp) -> ExecutionPath {
    match op {
        NetOp::Socket { .. }
        | NetOp::Connect { .. }
        | NetOp::Bind { .. }
        | NetOp::Listen { .. }
        | NetOp::Accept { .. }
        | NetOp::Send { .. }
        | NetOp::Recv { .. }
        | NetOp::Shutdown { .. }
        | NetOp::Close { .. } => ExecutionPath::IoUring,
        NetOp::SendTo { .. } | NetOp::RecvFrom { .. } => ExecutionPath::IoUring,
    }
}

pub async fn resolve_addrs<A>(addr: A) -> io::Result<Vec<SocketAddr>>
where
    A: ToSocketAddrs + Send + 'static,
{
    offload(move || {
        let addrs = addr.to_socket_addrs()?.collect::<Vec<_>>();
        if addrs.is_empty() {
            Err(io::Error::new(
                io::ErrorKind::InvalidInput,
                "address resolved to no socket addresses",
            ))
        } else {
            Ok(addrs)
        }
    })
    .await
}

pub async fn socket(op: NetOp) -> io::Result<OwnedFd> {
    let NetOp::Socket {
        domain,
        socket_type,
        protocol,
        flags,
    } = op
    else {
        unreachable!("socket backend called with non-socket op");
    };

    match submit_uring::<OwnedFd, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_SOCKET;
            sqe.fd = domain;
            sqe.off = socket_type as u64;
            sqe.len = protocol as u32;
            sqe.op_flags = flags;
        },
        move |cqe| cqe_to_result(cqe).map(|fd| unsafe { OwnedFd::from_raw_fd(fd as RawFd) }),
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            offload(move || socket_sync(domain, socket_type, protocol, flags)).await
        }
        result => result,
    }
}

pub async fn connect(op: NetOp) -> io::Result<()> {
    let NetOp::Connect { fd, addr } = op else {
        unreachable!("connect backend called with non-connect op");
    };

    let raw_addr = RawSocketAddr::from_socket_addr(addr);
    let fallback_addr = raw_addr;
    let addr_ptr = raw_addr.as_ptr();
    let addr_len = raw_addr.len();
    match submit_uring::<(), _>(
        move |sqe| {
            sqe.opcode = IORING_OP_CONNECT;
            sqe.fd = fd;
            sqe.addr = addr_ptr as u64;
            sqe.off = addr_len as u64;
        },
        move |cqe| {
            let _raw_addr = raw_addr;
            cqe_to_result(cqe).map(|_| ())
        },
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            offload(move || connect_sync(fd, fallback_addr)).await
        }
        result => result,
    }
}

pub async fn bind(op: NetOp) -> io::Result<()> {
    let NetOp::Bind { fd, addr } = op else {
        unreachable!("bind backend called with non-bind op");
    };

    let raw_addr = RawSocketAddr::from_socket_addr(addr);
    let fallback_addr = raw_addr;
    let addr_ptr = raw_addr.as_ptr();
    let addr_len = raw_addr.len();
    match submit_uring::<(), _>(
        move |sqe| {
            sqe.opcode = IORING_OP_BIND;
            sqe.fd = fd;
            sqe.addr = addr_ptr as u64;
            sqe.off = addr_len as u64;
        },
        move |cqe| {
            let _raw_addr = raw_addr;
            cqe_to_result(cqe).map(|_| ())
        },
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            offload(move || bind_sync(fd, fallback_addr)).await
        }
        result => result,
    }
}

pub async fn listen(op: NetOp) -> io::Result<()> {
    let NetOp::Listen { fd, backlog } = op else {
        unreachable!("listen backend called with non-listen op");
    };

    match submit_uring::<(), _>(
        move |sqe| {
            sqe.opcode = IORING_OP_LISTEN;
            sqe.fd = fd;
            sqe.len = backlog as u32;
        },
        move |cqe| cqe_to_result(cqe).map(|_| ()),
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            offload(move || listen_sync(fd, backlog)).await
        }
        result => result,
    }
}

pub async fn accept(op: NetOp) -> io::Result<AcceptedSocket> {
    let NetOp::Accept { fd } = op else {
        unreachable!("accept backend called with non-accept op");
    };

    let mut storage = Box::new(MaybeUninit::<libc::sockaddr_storage>::zeroed());
    let mut addr_len = Box::new(std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t);
    let storage_ptr = storage.as_mut_ptr();
    let addr_len_ptr = addr_len.as_mut() as *mut libc::socklen_t;

    match submit_uring::<AcceptedSocket, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_ACCEPT;
            sqe.fd = fd;
            sqe.addr = storage_ptr as u64;
            sqe.off = addr_len_ptr as u64;
        },
        move |cqe| {
            let accepted_fd = cqe_to_result(cqe)? as RawFd;
            let storage = unsafe { storage.assume_init() };
            let peer_addr = socket_addr_from_storage(&storage, *addr_len)?;
            Ok(AcceptedSocket {
                fd: accepted_fd,
                peer_addr,
            })
        },
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => offload(move || accept_sync(fd)).await,
        result => result,
    }
}

pub async fn send(op: NetOp) -> io::Result<usize> {
    let NetOp::Send { fd, data, flags } = op else {
        unreachable!("send backend called with non-send op");
    };

    // Capability known: io_uring SEND is not supported — skip to offload directly.
    if SEND_URING_SUPPORTED.with(|c| c.get()) == Some(false) {
        return offload(move || send_sync(fd, data, flags)).await;
    }

    // Capability known: io_uring SEND works — submit without a fallback clone.
    if SEND_URING_SUPPORTED.with(|c| c.get()) == Some(true) {
        let data_ptr = data.as_ptr();
        let data_len = data.len();
        return submit_uring::<usize, _>(
            move |sqe| {
                sqe.opcode = IORING_OP_SEND;
                sqe.fd = fd;
                sqe.addr = data_ptr as u64;
                sqe.len = data_len as u32;
                sqe.op_flags = flags as u32;
            },
            move |cqe| {
                let _data = data;
                cqe_to_result(cqe).map(|written| written as usize)
            },
        )
        .await;
    }

    // Capability unknown: probe with a one-time clone. Cache the result.
    let fallback_data = data.clone();
    let data_ptr = data.as_ptr();
    let data_len = data.len();
    match submit_uring::<usize, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_SEND;
            sqe.fd = fd;
            sqe.addr = data_ptr as u64;
            sqe.len = data_len as u32;
            sqe.op_flags = flags as u32;
        },
        move |cqe| {
            let _data = data;
            cqe_to_result(cqe).map(|written| written as usize)
        },
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            SEND_URING_SUPPORTED.with(|c| c.set(Some(false)));
            offload(move || send_sync(fd, fallback_data, flags)).await
        }
        result => {
            if result.is_ok() {
                SEND_URING_SUPPORTED.with(|c| c.set(Some(true)));
            }
            result
        }
    }
}

pub async fn send_to(op: NetOp) -> io::Result<usize> {
    let NetOp::SendTo {
        fd,
        target,
        data,
        flags,
    } = op
    else {
        unreachable!("send_to backend called with non-send_to op");
    };

    let raw_addr = Box::new(RawSocketAddr::from_socket_addr(target));
    let mut iov = Box::new(libc::iovec {
        iov_base: data.as_ptr() as *mut c_void,
        iov_len: data.len(),
    });
    let mut msg = Box::new(unsafe { std::mem::zeroed::<libc::msghdr>() });
    msg.msg_name = raw_addr.as_ptr() as *mut c_void;
    msg.msg_namelen = raw_addr.len();
    msg.msg_iov = iov.as_mut() as *mut libc::iovec;
    msg.msg_iovlen = 1;
    let msg_ptr = msg.as_mut() as *mut libc::msghdr as u64;
    let iov = SendIovec(iov);
    let msg = SendMsghdr(msg);

    submit_uring::<usize, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_SENDMSG;
            sqe.fd = fd;
            sqe.addr = msg_ptr;
            sqe.op_flags = flags as u32;
        },
        move |cqe| {
            let _raw_addr = raw_addr;
            let _iov = iov;
            let _msg = msg;
            let _data = data;
            cqe_to_result(cqe).map(|written| written as usize)
        },
    )
    .await
}

pub async fn recv(op: NetOp) -> io::Result<Vec<u8>> {
    let NetOp::Recv { fd, len, flags } = op else {
        unreachable!("recv backend called with non-recv op");
    };

    let mut buffer = vec![0; len];
    let buffer_ptr = buffer.as_mut_ptr();
    let buffer_len = buffer.len();
    match submit_uring::<Vec<u8>, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_RECV;
            sqe.fd = fd;
            sqe.addr = buffer_ptr as u64;
            sqe.len = buffer_len as u32;
            sqe.op_flags = flags as u32;
        },
        move |cqe| {
            let read = cqe_to_result(cqe)? as usize;
            buffer.truncate(read);
            Ok(buffer)
        },
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            offload(move || recv_sync(fd, len, flags)).await
        }
        result => result,
    }
}

pub async fn recv_from(op: NetOp) -> io::Result<ReceivedDatagram> {
    let NetOp::RecvFrom { fd, len, flags } = op else {
        unreachable!("recv_from backend called with non-recv_from op");
    };

    let mut data = vec![0u8; len];
    let mut storage = Box::new(MaybeUninit::<libc::sockaddr_storage>::zeroed());
    let mut iov = Box::new(libc::iovec {
        iov_base: data.as_mut_ptr() as *mut c_void,
        iov_len: data.len(),
    });
    let mut msg = Box::new(unsafe { std::mem::zeroed::<libc::msghdr>() });
    msg.msg_name = storage.as_mut_ptr() as *mut c_void;
    msg.msg_namelen = std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
    msg.msg_iov = iov.as_mut() as *mut libc::iovec;
    msg.msg_iovlen = 1;
    let msg_ptr = msg.as_mut() as *mut libc::msghdr as u64;
    let iov = SendIovec(iov);
    let msg = SendMsghdr(msg);

    match submit_uring::<ReceivedDatagram, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_RECVMSG;
            sqe.fd = fd;
            sqe.addr = msg_ptr;
            sqe.op_flags = flags as u32;
        },
        move |cqe| {
            let _iov = iov;
            let addr_len = msg.0.msg_namelen;
            drop(msg);
            let read = cqe_to_result(cqe)? as usize;
            data.truncate(read);
            let storage = unsafe { storage.assume_init() };
            let peer_addr = socket_addr_from_storage(&storage, addr_len)?;
            Ok(ReceivedDatagram { data, peer_addr })
        },
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            offload(move || recv_from_sync(fd, len, flags)).await
        }
        result => result,
    }
}

pub async fn shutdown(op: NetOp) -> io::Result<()> {
    let NetOp::Shutdown { fd, how } = op else {
        unreachable!("shutdown backend called with non-shutdown op");
    };

    let fallback_how = how;
    match submit_uring::<(), _>(
        move |sqe| {
            sqe.opcode = IORING_OP_SHUTDOWN;
            sqe.fd = fd;
            sqe.len = shutdown_how(how) as u32;
        },
        move |cqe| cqe_to_result(cqe).map(|_| ()),
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => {
            offload(move || shutdown_sync(fd, fallback_how)).await
        }
        result => result,
    }
}

pub async fn close(op: NetOp) -> io::Result<()> {
    let NetOp::Close { fd } = op else {
        unreachable!("close backend called with non-close op");
    };

    match submit_uring::<(), _>(
        move |sqe| {
            sqe.opcode = IORING_OP_CLOSE;
            sqe.fd = fd;
        },
        move |cqe| cqe_to_result(cqe).map(|_| ()),
    )
    .await
    {
        Err(error) if should_fallback_to_offload(&error) => offload(move || close_sync(fd)).await,
        result => result,
    }
}

pub async fn connect_stream(addr: SocketAddr) -> io::Result<OwnedFd> {
    let socket = socket(NetOp::Socket {
        domain: socket_domain(addr),
        socket_type: libc::SOCK_STREAM,
        protocol: 0,
        flags: libc::SOCK_CLOEXEC as u32,
    })
    .await?;

    let connect_result = connect(NetOp::Connect {
        fd: socket.as_raw_fd(),
        addr,
    })
    .await;
    match connect_result {
        Ok(()) => Ok(socket),
        Err(error) => Err(error),
    }
}

pub async fn bind_listener(addr: SocketAddr, backlog: Option<i32>) -> io::Result<OwnedFd> {
    let listener = socket(NetOp::Socket {
        domain: socket_domain(addr),
        socket_type: libc::SOCK_STREAM,
        protocol: 0,
        flags: libc::SOCK_CLOEXEC as u32,
    })
    .await?;

    set_reuse_addr(listener.as_raw_fd(), true)?;

    bind(NetOp::Bind {
        fd: listener.as_raw_fd(),
        addr,
    })
    .await?;
    listen(NetOp::Listen {
        fd: listener.as_raw_fd(),
        backlog: backlog.unwrap_or(DEFAULT_LISTENER_BACKLOG),
    })
    .await?;
    Ok(listener)
}

pub async fn bind_datagram(addr: SocketAddr) -> io::Result<OwnedFd> {
    let socket = socket(NetOp::Socket {
        domain: socket_domain(addr),
        socket_type: libc::SOCK_DGRAM,
        protocol: 0,
        flags: libc::SOCK_CLOEXEC as u32,
    })
    .await?;

    bind(NetOp::Bind {
        fd: socket.as_raw_fd(),
        addr,
    })
    .await?;
    Ok(socket)
}

pub async fn duplicate(fd: RawFd) -> io::Result<OwnedFd> {
    offload(move || {
        let duplicated = cvt(unsafe { libc::fcntl(fd, libc::F_DUPFD_CLOEXEC, 0) })?;
        Ok(unsafe { OwnedFd::from_raw_fd(duplicated) })
    })
    .await
}

pub async fn recv_timeout(
    fd: RawFd,
    len: usize,
    flags: i32,
    timeout: Duration,
) -> io::Result<Vec<u8>> {
    let mut buffer = vec![0u8; len];
    let buffer_ptr = buffer.as_mut_ptr();
    let buffer_len = buffer.len();
    submit_uring_with_linked_timeout::<Vec<u8>, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_RECV;
            sqe.fd = fd;
            sqe.addr = buffer_ptr as u64;
            sqe.len = buffer_len as u32;
            sqe.op_flags = flags as u32;
        },
        timeout,
        move |cqe| {
            let read = cqe_to_timed_result(cqe)? as usize;
            buffer.truncate(read);
            Ok(buffer)
        },
    )
    .await
}

pub async fn send_timeout(
    fd: RawFd,
    data: Vec<u8>,
    flags: i32,
    timeout: Duration,
) -> io::Result<usize> {
    let data_ptr = data.as_ptr();
    let data_len = data.len();
    submit_uring_with_linked_timeout::<usize, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_SEND;
            sqe.fd = fd;
            sqe.addr = data_ptr as u64;
            sqe.len = data_len as u32;
            sqe.op_flags = flags as u32;
        },
        timeout,
        move |cqe| {
            let _data = data;
            cqe_to_timed_result(cqe).map(|written| written as usize)
        },
    )
    .await
}

pub async fn recv_from_timeout(
    fd: RawFd,
    len: usize,
    flags: i32,
    timeout: Duration,
) -> io::Result<ReceivedDatagram> {
    let mut data = vec![0u8; len];
    let mut storage = Box::new(MaybeUninit::<libc::sockaddr_storage>::zeroed());
    let mut iov = Box::new(libc::iovec {
        iov_base: data.as_mut_ptr() as *mut c_void,
        iov_len: data.len(),
    });
    let mut msg = Box::new(unsafe { std::mem::zeroed::<libc::msghdr>() });
    msg.msg_name = storage.as_mut_ptr() as *mut c_void;
    msg.msg_namelen = std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
    msg.msg_iov = iov.as_mut() as *mut libc::iovec;
    msg.msg_iovlen = 1;
    let msg_ptr = msg.as_mut() as *mut libc::msghdr as u64;
    let iov = SendIovec(iov);
    let msg = SendMsghdr(msg);

    submit_uring_with_linked_timeout::<ReceivedDatagram, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_RECVMSG;
            sqe.fd = fd;
            sqe.addr = msg_ptr;
            sqe.op_flags = flags as u32;
        },
        timeout,
        move |cqe| {
            let _iov = iov;
            let addr_len = msg.0.msg_namelen;
            drop(msg);
            let read = cqe_to_timed_result(cqe)? as usize;
            data.truncate(read);
            let storage = unsafe { storage.assume_init() };
            let peer_addr = socket_addr_from_storage(&storage, addr_len)?;
            Ok(ReceivedDatagram { data, peer_addr })
        },
    )
    .await
}

pub async fn send_to_timeout(
    fd: RawFd,
    data: Vec<u8>,
    target: SocketAddr,
    flags: i32,
    timeout: Duration,
) -> io::Result<usize> {
    let raw_addr = Box::new(RawSocketAddr::from_socket_addr(target));
    let mut iov = Box::new(libc::iovec {
        iov_base: data.as_ptr() as *mut c_void,
        iov_len: data.len(),
    });
    let mut msg = Box::new(unsafe { std::mem::zeroed::<libc::msghdr>() });
    msg.msg_name = raw_addr.as_ptr() as *mut c_void;
    msg.msg_namelen = raw_addr.len();
    msg.msg_iov = iov.as_mut() as *mut libc::iovec;
    msg.msg_iovlen = 1;
    let msg_ptr = msg.as_mut() as *mut libc::msghdr as u64;
    let iov = SendIovec(iov);
    let msg = SendMsghdr(msg);

    submit_uring_with_linked_timeout::<usize, _>(
        move |sqe| {
            sqe.opcode = IORING_OP_SENDMSG;
            sqe.fd = fd;
            sqe.addr = msg_ptr;
            sqe.op_flags = flags as u32;
        },
        timeout,
        move |cqe| {
            let _raw_addr = raw_addr;
            let _iov = iov;
            let _msg = msg;
            let _data = data;
            cqe_to_timed_result(cqe).map(|written| written as usize)
        },
    )
    .await
}

pub async fn connect_stream_timeout(addr: SocketAddr, timeout: Duration) -> io::Result<OwnedFd> {
    let socket = socket(NetOp::Socket {
        domain: socket_domain(addr),
        socket_type: libc::SOCK_STREAM,
        protocol: 0,
        flags: libc::SOCK_CLOEXEC as u32,
    })
    .await?;

    let fd = socket.as_raw_fd();
    let raw_addr = RawSocketAddr::from_socket_addr(addr);
    let addr_ptr = raw_addr.as_ptr();
    let addr_len = raw_addr.len();

    submit_uring_with_linked_timeout::<(), _>(
        move |sqe| {
            sqe.opcode = IORING_OP_CONNECT;
            sqe.fd = fd;
            sqe.addr = addr_ptr as u64;
            sqe.off = addr_len as u64;
        },
        timeout,
        move |cqe| {
            let _raw_addr = raw_addr;
            cqe_to_timed_result(cqe).map(|_| ())
        },
    )
    .await?;

    Ok(socket)
}

pub fn local_addr(fd: RawFd) -> io::Result<SocketAddr> {
    socket_addr_with(libc::getsockname, fd)
}

pub fn peer_addr(fd: RawFd) -> io::Result<SocketAddr> {
    socket_addr_with(libc::getpeername, fd)
}

pub fn nodelay(fd: RawFd) -> io::Result<bool> {
    let mut value = 0;
    let mut len = std::mem::size_of::<libc::c_int>() as libc::socklen_t;
    cvt(unsafe {
        libc::getsockopt(
            fd,
            libc::IPPROTO_TCP,
            libc::TCP_NODELAY,
            &mut value as *mut libc::c_int as *mut c_void,
            &mut len,
        )
    })?;
    Ok(value != 0)
}

pub fn broadcast(fd: RawFd) -> io::Result<bool> {
    getsockopt_int(fd, libc::SOL_SOCKET, libc::SO_BROADCAST).map(|value| value != 0)
}

pub fn set_broadcast(fd: RawFd, enabled: bool) -> io::Result<()> {
    setsockopt_int(fd, libc::SOL_SOCKET, libc::SO_BROADCAST, enabled.into())
}

pub fn ttl(fd: RawFd) -> io::Result<u32> {
    match socket_family(fd)? {
        libc::AF_INET => {
            getsockopt_int(fd, libc::IPPROTO_IP, libc::IP_TTL).map(|value| value as u32)
        }
        libc::AF_INET6 => getsockopt_int(fd, libc::IPPROTO_IPV6, libc::IPV6_UNICAST_HOPS)
            .map(|value| value as u32),
        family => Err(io::Error::new(
            io::ErrorKind::InvalidInput,
            format!("unsupported socket family {family} for TTL"),
        )),
    }
}

pub fn set_ttl(fd: RawFd, ttl: u32) -> io::Result<()> {
    let ttl = i32::try_from(ttl)
        .map_err(|_| io::Error::new(io::ErrorKind::InvalidInput, "TTL exceeds i32 range"))?;
    match socket_family(fd)? {
        libc::AF_INET => setsockopt_int(fd, libc::IPPROTO_IP, libc::IP_TTL, ttl),
        libc::AF_INET6 => setsockopt_int(fd, libc::IPPROTO_IPV6, libc::IPV6_UNICAST_HOPS, ttl),
        family => Err(io::Error::new(
            io::ErrorKind::InvalidInput,
            format!("unsupported socket family {family} for TTL"),
        )),
    }
}

pub fn set_nodelay(fd: RawFd, enabled: bool) -> io::Result<()> {
    let value: libc::c_int = enabled.into();
    cvt(unsafe {
        libc::setsockopt(
            fd,
            libc::IPPROTO_TCP,
            libc::TCP_NODELAY,
            &value as *const libc::c_int as *const c_void,
            std::mem::size_of_val(&value) as libc::socklen_t,
        )
    })
    .map(|_| ())
}

pub type RecvFuture = Pin<Box<dyn Future<Output = io::Result<Vec<u8>>> + 'static>>;
pub type SendFuture = Pin<Box<dyn Future<Output = io::Result<usize>> + 'static>>;
pub type ShutdownFuture = Pin<Box<dyn Future<Output = io::Result<()>> + 'static>>;

pub fn recv_future(fd: RawFd, len: usize) -> RecvFuture {
    Box::pin(recv(NetOp::Recv { fd, len, flags: 0 }))
}

pub fn send_future(fd: RawFd, data: Vec<u8>) -> SendFuture {
    Box::pin(send(NetOp::Send { fd, data, flags: 0 }))
}

pub fn shutdown_future(fd: RawFd, how: Shutdown) -> ShutdownFuture {
    Box::pin(shutdown(NetOp::Shutdown { fd, how }))
}

async fn submit_uring<T: Send + 'static, M>(
    fill: impl FnOnce(&mut IoUringSqe),
    map: M,
) -> io::Result<T>
where
    M: FnOnce(IoUringCqe) -> io::Result<T> + Send + 'static,
{
    let (future, handle) = completion_for_current_thread::<io::Result<T>>();
    let callback_handle = handle.clone();
    let token = with_current_driver(|driver| {
        driver.submit_operation(fill, move |cqe| {
            callback_handle.complete(map(cqe));
        })
    })?;

    handle.set_cancel(move || {
        let _ = with_current_driver(|driver| driver.cancel_operation(token));
    });

    future.await
}

/// Like [`submit_uring`] but pairs the main SQE with an `IORING_OP_LINK_TIMEOUT`.
///
/// If the timeout elapses before the main op completes, the completion callback
/// receives a CQE with `res = -ECANCELED`.  Callers should map `-ECANCELED` to
/// `io::ErrorKind::TimedOut`.
async fn submit_uring_with_linked_timeout<T: Send + 'static, M>(
    fill: impl FnOnce(&mut IoUringSqe),
    timeout: Duration,
    map: M,
) -> io::Result<T>
where
    M: FnOnce(IoUringCqe) -> io::Result<T> + Send + 'static,
{
    let (future, handle) = completion_for_current_thread::<io::Result<T>>();
    let callback_handle = handle.clone();
    let token = with_current_driver(|driver| {
        driver.submit_operation_with_linked_timeout(fill, timeout, move |cqe| {
            callback_handle.complete(map(cqe));
        })
    })?;

    handle.set_cancel(move || {
        let _ = with_current_driver(|driver| driver.cancel_operation(token));
    });

    future.await
}

async fn offload<T: Send + 'static>(
    task: impl FnOnce() -> io::Result<T> + Send + 'static,
) -> io::Result<T> {
    let (future, handle) = completion_for_current_thread::<io::Result<T>>();
    thread::Builder::new()
        .name("ruin-runtime-net-offload".into())
        .spawn(move || handle.complete(task()))
        .map_err(io::Error::other)?;
    future.await
}

fn socket_domain(addr: SocketAddr) -> i32 {
    match addr {
        SocketAddr::V4(_) => libc::AF_INET,
        SocketAddr::V6(_) => libc::AF_INET6,
    }
}

fn shutdown_how(how: Shutdown) -> i32 {
    match how {
        Shutdown::Read => libc::SHUT_RD,
        Shutdown::Write => libc::SHUT_WR,
        Shutdown::Both => libc::SHUT_RDWR,
    }
}

fn socket_addr_with(
    op: unsafe extern "C" fn(RawFd, *mut libc::sockaddr, *mut libc::socklen_t) -> libc::c_int,
    fd: RawFd,
) -> io::Result<SocketAddr> {
    let mut storage = MaybeUninit::<libc::sockaddr_storage>::zeroed();
    let mut len = std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
    cvt(unsafe { op(fd, storage.as_mut_ptr().cast::<libc::sockaddr>(), &mut len) })?;
    let storage = unsafe { storage.assume_init() };
    socket_addr_from_storage(&storage, len)
}

fn set_reuse_addr(fd: RawFd, enabled: bool) -> io::Result<()> {
    setsockopt_int(fd, libc::SOL_SOCKET, libc::SO_REUSEADDR, enabled.into())
}

fn socket_family(fd: RawFd) -> io::Result<i32> {
    let mut storage = MaybeUninit::<libc::sockaddr_storage>::zeroed();
    let mut len = std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
    cvt(unsafe { libc::getsockname(fd, storage.as_mut_ptr().cast::<libc::sockaddr>(), &mut len) })?;
    let storage = unsafe { storage.assume_init() };
    Ok(storage.ss_family as i32)
}

fn getsockopt_int(fd: RawFd, level: i32, name: i32) -> io::Result<i32> {
    let mut value = 0;
    let mut len = std::mem::size_of::<libc::c_int>() as libc::socklen_t;
    cvt(unsafe {
        libc::getsockopt(
            fd,
            level,
            name,
            &mut value as *mut libc::c_int as *mut c_void,
            &mut len,
        )
    })?;
    Ok(value)
}

fn setsockopt_int(fd: RawFd, level: i32, name: i32, value: i32) -> io::Result<()> {
    cvt(unsafe {
        libc::setsockopt(
            fd,
            level,
            name,
            &value as *const libc::c_int as *const c_void,
            std::mem::size_of_val(&value) as libc::socklen_t,
        )
    })
    .map(|_| ())
}

fn socket_addr_from_storage(
    storage: &libc::sockaddr_storage,
    len: libc::socklen_t,
) -> io::Result<SocketAddr> {
    match storage.ss_family as i32 {
        libc::AF_INET => {
            if len < std::mem::size_of::<libc::sockaddr_in>() as libc::socklen_t {
                return Err(io::Error::new(
                    io::ErrorKind::InvalidData,
                    "short IPv4 socket address from kernel",
                ));
            }

            let addr = unsafe { *(storage as *const _ as *const libc::sockaddr_in) };
            Ok(SocketAddr::V4(SocketAddrV4::new(
                Ipv4Addr::from(addr.sin_addr.s_addr.to_ne_bytes()),
                u16::from_be(addr.sin_port),
            )))
        }
        libc::AF_INET6 => {
            if len < std::mem::size_of::<libc::sockaddr_in6>() as libc::socklen_t {
                return Err(io::Error::new(
                    io::ErrorKind::InvalidData,
                    "short IPv6 socket address from kernel",
                ));
            }

            let addr = unsafe { *(storage as *const _ as *const libc::sockaddr_in6) };
            Ok(SocketAddr::V6(SocketAddrV6::new(
                Ipv6Addr::from(addr.sin6_addr.s6_addr),
                u16::from_be(addr.sin6_port),
                addr.sin6_flowinfo,
                addr.sin6_scope_id,
            )))
        }
        family => Err(io::Error::new(
            io::ErrorKind::InvalidData,
            format!("unsupported socket address family {family}"),
        )),
    }
}

#[derive(Clone, Copy)]
struct RawSocketAddr {
    storage: libc::sockaddr_storage,
    len: libc::socklen_t,
}

impl RawSocketAddr {
    fn from_socket_addr(addr: SocketAddr) -> Self {
        match addr {
            SocketAddr::V4(addr) => {
                let sockaddr = libc::sockaddr_in {
                    sin_family: libc::AF_INET as libc::sa_family_t,
                    sin_port: addr.port().to_be(),
                    sin_addr: libc::in_addr {
                        s_addr: u32::from_ne_bytes(addr.ip().octets()),
                    },
                    sin_zero: [0; 8],
                };
                let mut storage =
                    unsafe { MaybeUninit::<libc::sockaddr_storage>::zeroed().assume_init() };
                unsafe {
                    std::ptr::write(
                        &mut storage as *mut libc::sockaddr_storage as *mut libc::sockaddr_in,
                        sockaddr,
                    );
                }
                Self {
                    storage,
                    len: std::mem::size_of::<libc::sockaddr_in>() as libc::socklen_t,
                }
            }
            SocketAddr::V6(addr) => {
                let sockaddr = libc::sockaddr_in6 {
                    sin6_family: libc::AF_INET6 as libc::sa_family_t,
                    sin6_port: addr.port().to_be(),
                    sin6_flowinfo: addr.flowinfo(),
                    sin6_addr: libc::in6_addr {
                        s6_addr: addr.ip().octets(),
                    },
                    sin6_scope_id: addr.scope_id(),
                };
                let mut storage =
                    unsafe { MaybeUninit::<libc::sockaddr_storage>::zeroed().assume_init() };
                unsafe {
                    std::ptr::write(
                        &mut storage as *mut libc::sockaddr_storage as *mut libc::sockaddr_in6,
                        sockaddr,
                    );
                }
                Self {
                    storage,
                    len: std::mem::size_of::<libc::sockaddr_in6>() as libc::socklen_t,
                }
            }
        }
    }

    fn as_ptr(&self) -> *const libc::sockaddr {
        &self.storage as *const libc::sockaddr_storage as *const libc::sockaddr
    }

    fn len(&self) -> libc::socklen_t {
        self.len
    }
}

fn cqe_to_result(cqe: IoUringCqe) -> io::Result<i32> {
    if cqe.res < 0 {
        Err(io::Error::from_raw_os_error(-cqe.res))
    } else {
        Ok(cqe.res)
    }
}

/// Like [`cqe_to_result`] but maps `-ECANCELED` (timeout fired) to `TimedOut`.
fn cqe_to_timed_result(cqe: IoUringCqe) -> io::Result<i32> {
    if cqe.res == -libc::ECANCELED {
        return Err(io::Error::new(
            io::ErrorKind::TimedOut,
            "socket operation timed out",
        ));
    }
    cqe_to_result(cqe)
}

fn cvt(value: libc::c_int) -> io::Result<libc::c_int> {
    if value == -1 {
        Err(io::Error::last_os_error())
    } else {
        Ok(value)
    }
}

fn should_fallback_to_offload(error: &io::Error) -> bool {
    matches!(
        error.raw_os_error(),
        Some(libc::EINVAL | libc::ENOSYS | libc::EOPNOTSUPP)
    )
}

fn socket_sync(domain: i32, socket_type: i32, protocol: i32, flags: u32) -> io::Result<OwnedFd> {
    let fd = cvt(unsafe { libc::socket(domain, socket_type | flags as i32, protocol) })?;
    Ok(unsafe { OwnedFd::from_raw_fd(fd) })
}

fn connect_sync(fd: RawFd, addr: RawSocketAddr) -> io::Result<()> {
    cvt(unsafe { libc::connect(fd, addr.as_ptr(), addr.len()) }).map(|_| ())
}

fn bind_sync(fd: RawFd, addr: RawSocketAddr) -> io::Result<()> {
    cvt(unsafe { libc::bind(fd, addr.as_ptr(), addr.len()) }).map(|_| ())
}

fn listen_sync(fd: RawFd, backlog: i32) -> io::Result<()> {
    cvt(unsafe { libc::listen(fd, backlog) }).map(|_| ())
}

fn accept_sync(fd: RawFd) -> io::Result<AcceptedSocket> {
    let mut storage = MaybeUninit::<libc::sockaddr_storage>::zeroed();
    let mut len = std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
    let accepted_fd = cvt(unsafe {
        libc::accept4(
            fd,
            storage.as_mut_ptr().cast::<libc::sockaddr>(),
            &mut len,
            libc::SOCK_CLOEXEC,
        )
    })?;
    let storage = unsafe { storage.assume_init() };
    let peer_addr = socket_addr_from_storage(&storage, len)?;
    Ok(AcceptedSocket {
        fd: accepted_fd,
        peer_addr,
    })
}

fn send_sync(fd: RawFd, data: Vec<u8>, flags: i32) -> io::Result<usize> {
    let written = unsafe { libc::send(fd, data.as_ptr().cast::<c_void>(), data.len(), flags) };
    cvt_long(written).map(|written| written as usize)
}

fn recv_sync(fd: RawFd, len: usize, flags: i32) -> io::Result<Vec<u8>> {
    let mut buffer = vec![0; len];
    let read = unsafe {
        libc::recv(
            fd,
            buffer.as_mut_ptr().cast::<c_void>(),
            buffer.len(),
            flags,
        )
    };
    let read = cvt_long(read)? as usize;
    buffer.truncate(read);
    Ok(buffer)
}

fn recv_from_sync(fd: RawFd, len: usize, flags: i32) -> io::Result<ReceivedDatagram> {
    let mut buffer = vec![0; len];
    let mut storage = MaybeUninit::<libc::sockaddr_storage>::zeroed();
    let mut addr_len = std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
    let read = unsafe {
        libc::recvfrom(
            fd,
            buffer.as_mut_ptr().cast::<c_void>(),
            buffer.len(),
            flags,
            storage.as_mut_ptr().cast::<libc::sockaddr>(),
            &mut addr_len,
        )
    };
    let read = cvt_long(read)? as usize;
    buffer.truncate(read);
    let storage = unsafe { storage.assume_init() };
    let peer_addr = socket_addr_from_storage(&storage, addr_len)?;
    Ok(ReceivedDatagram {
        data: buffer,
        peer_addr,
    })
}

fn shutdown_sync(fd: RawFd, how: Shutdown) -> io::Result<()> {
    cvt(unsafe { libc::shutdown(fd, shutdown_how(how)) }).map(|_| ())
}

fn close_sync(fd: RawFd) -> io::Result<()> {
    cvt(unsafe { libc::close(fd) }).map(|_| ())
}

/// Wrapper making `Box<libc::iovec>` sendable across the async CQE boundary.
///
/// Safety: `iov_base` points into a `Vec<u8>` that is owned by the same
/// closure, so the pointer is valid until the CQE fires and the closure drops.
struct SendIovec(#[allow(dead_code)] Box<libc::iovec>);
unsafe impl Send for SendIovec {}

/// Wrapper making `Box<libc::msghdr>` sendable across the async CQE boundary.
///
/// Safety: all raw pointers inside the `msghdr` (`msg_name`, `msg_iov`) point
/// into heap storage owned by the same closure, so they are valid until the
/// CQE fires and the closure drops.
struct SendMsghdr(Box<libc::msghdr>);
unsafe impl Send for SendMsghdr {}

fn cvt_long(value: libc::ssize_t) -> io::Result<libc::ssize_t> {
    if value == -1 {
        Err(io::Error::last_os_error())
    } else {
        Ok(value)
    }
}