diff --git a/proxy/src/connection.rs b/proxy/src/connection.rs
index 01d461b23..9d6ebfbd3 100644
--- a/proxy/src/connection.rs
+++ b/proxy/src/connection.rs
@@ -3,15 +3,14 @@ use futures::*;
 use std;
 use std::io;
 use std::net::SocketAddr;
-use tokio_core;
-use tokio_core::net::{TcpListener, TcpStreamNew};
+use tokio_core::net::{TcpListener, TcpStreamNew, TcpStream};
 use tokio_core::reactor::Handle;
 use tokio_io::{AsyncRead, AsyncWrite};
 
 use config::Addr;
 use transport::GetOriginalDst;
 
-pub type PlaintextSocket = tokio_core::net::TcpStream;
+pub type PlaintextSocket = TcpStream;
 
 pub struct BoundPort {
     inner: std::net::TcpListener,
@@ -165,10 +164,20 @@ impl io::Write for Connection {
 
 impl AsyncWrite for Connection {
     fn shutdown(&mut self) -> Poll<(), io::Error> {
+        use std::net::Shutdown;
         use self::Connection::*;
 
         match *self {
-            Plain(ref mut t) => t.shutdown(),
+            Plain(ref mut t) => {
+                try_ready!(AsyncWrite::shutdown(t));
+                // TCP shutdown the write side.
+                //
+                // If we're shutting down, then we definitely won't write
+                // anymore. So, we should tell the remote about this. This
+                // is relied upon in our TCP proxy, to start shutting down
+                // the pipe if one side closes.
+                TcpStream::shutdown(t, Shutdown::Write).map(Async::Ready)
+            },
         }
     }
 
diff --git a/proxy/src/transparency/tcp.rs b/proxy/src/transparency/tcp.rs
index 016c83cbd..3ab74b871 100644
--- a/proxy/src/transparency/tcp.rs
+++ b/proxy/src/transparency/tcp.rs
@@ -1,11 +1,12 @@
+use std::io;
 use std::sync::Arc;
 use std::time::Duration;
 
-use futures::{future, Future};
+use bytes::{Buf, BufMut};
+use futures::{future, Async, Future, Poll};
 use tokio_connect::Connect;
 use tokio_core::reactor::Handle;
 use tokio_io::{AsyncRead, AsyncWrite};
-use tokio_io::io::copy;
 
 use conduit_proxy_controller_grpc::common;
 use ctx::transport::{Client as ClientCtx, Server as ServerCtx};
@@ -71,14 +72,188 @@ impl Proxy {
         let fut = connect.connect()
             .map_err(|e| debug!("tcp connect error: {:?}", e))
             .and_then(move |tcp_out| {
-                let (in_r, in_w) = tcp_in.split();
-                let (out_r, out_w) = tcp_out.split();
-
-                copy(in_r, out_w)
-                    .join(copy(out_r, in_w))
-                    .map(|_| ())
+                Duplex::new(tcp_in, tcp_out)
                     .map_err(|e| debug!("tcp error: {}", e))
             });
         Box::new(fut)
     }
 }
+
+/// A future piping data bi-directionally to In and Out.
+struct Duplex<In, Out> {
+    half_in: HalfDuplex<In>,
+    half_out: HalfDuplex<Out>,
+}
+
+struct HalfDuplex<T> {
+    // None means socket met eof, and bytes have been drained into other half.
+    buf: Option<CopyBuf>,
+    is_shutdown: bool,
+    io: T,
+}
+
+/// A buffer used to copy bytes from one IO to another.
+///
+/// Keeps read and write positions.
+struct CopyBuf {
+    // TODO:
+    // In linkerd-tcp, a shared buffer is used to start, and an allocation is
+    // only made if NotReady is found trying to flush the buffer. We could
+    // consider making the same optimization here.
+    buf: Box<[u8]>,
+    read_pos: usize,
+    write_pos: usize,
+}
+
+impl<In, Out> Duplex<In, Out>
+where
+    In: AsyncRead + AsyncWrite,
+    Out: AsyncRead + AsyncWrite,
+{
+    fn new(in_io: In, out_io: Out) -> Self {
+        Duplex {
+            half_in: HalfDuplex::new(in_io),
+            half_out: HalfDuplex::new(out_io),
+        }
+    }
+}
+
+impl<In, Out> Future for Duplex<In, Out>
+where
+    In: AsyncRead + AsyncWrite,
+    Out: AsyncRead + AsyncWrite,
+{
+    type Item = ();
+    type Error = io::Error;
+
+    fn poll(&mut self) -> Poll<Self::Item, Self::Error> {
+        // This purposefully ignores the Async part, since we don't want to
+        // return early if the first half isn't ready, but the other half
+        // could make progress.
+        self.half_in.copy_into(&mut self.half_out)?;
+        self.half_out.copy_into(&mut self.half_in)?;
+
+        if self.half_in.is_done() && self.half_out.is_done() {
+            Ok(Async::Ready(()))
+        } else {
+            Ok(Async::NotReady)
+        }
+    }
+}
+
+impl<T> HalfDuplex<T>
+where
+    T: AsyncRead,
+{
+    fn new(io: T) -> Self {
+        Self {
+            buf: Some(CopyBuf::new()),
+            is_shutdown: false,
+            io,
+        }
+    }
+
+    fn copy_into<U>(&mut self, dst: &mut HalfDuplex<U>) -> Poll<(), io::Error>
+    where
+        U: AsyncWrite,
+    {
+        loop {
+            try_ready!(self.read());
+            try_ready!(self.write_into(dst));
+
+            if self.buf.is_none() && !dst.is_shutdown {
+                try_ready!(dst.io.shutdown());
+                dst.is_shutdown = true;
+
+                return Ok(Async::Ready(()));
+            }
+        }
+    }
+
+    fn read(&mut self) -> Poll<(), io::Error> {
+        let mut is_eof = false;
+        if let Some(ref mut buf) = self.buf {
+            if !buf.has_remaining() {
+                buf.reset();
+                let n = try_ready!(self.io.read_buf(buf));
+                is_eof = n == 0;
+            }
+        }
+
+        if is_eof {
+            self.buf.take();
+        }
+
+        Ok(Async::Ready(()))
+    }
+
+    fn write_into<U>(&mut self, dst: &mut HalfDuplex<U>) -> Poll<(), io::Error>
+    where
+        U: AsyncWrite,
+    {
+        if let Some(ref mut buf) = self.buf {
+            while buf.has_remaining() {
+                let n = try_ready!(dst.io.write_buf(buf));
+                if n == 0 {
+                    return Err(write_zero());
+                }
+            }
+        }
+
+        Ok(Async::Ready(()))
+    }
+
+    fn is_done(&self) -> bool {
+        self.is_shutdown
+    }
+}
+
+fn write_zero() -> io::Error {
+    io::Error::new(io::ErrorKind::WriteZero, "write zero bytes")
+}
+
+impl CopyBuf {
+    fn new() -> Self {
+        CopyBuf {
+            buf: Box::new([0; 4096]),
+            read_pos: 0,
+            write_pos: 0,
+        }
+    }
+
+    fn reset(&mut self) {
+        debug_assert_eq!(self.read_pos, self.write_pos);
+        self.read_pos = 0;
+        self.write_pos = 0;
+    }
+}
+
+impl Buf for CopyBuf {
+    fn remaining(&self) -> usize {
+        self.write_pos - self.read_pos
+    }
+
+    fn bytes(&self) -> &[u8] {
+        &self.buf[self.read_pos..self.write_pos]
+    }
+
+    fn advance(&mut self, cnt: usize) {
+        assert!(self.write_pos >= self.read_pos + cnt);
+        self.read_pos += cnt;
+    }
+}
+
+impl BufMut for CopyBuf {
+    fn remaining_mut(&self) -> usize {
+        self.buf.len() - self.write_pos
+    }
+
+    unsafe fn bytes_mut(&mut self) -> &mut [u8] {
+        &mut self.buf[self.write_pos..]
+    }
+
+    unsafe fn advance_mut(&mut self, cnt: usize) {
+        assert!(self.buf.len() >= self.write_pos + cnt);
+        self.write_pos += cnt;
+    }
+}
diff --git a/proxy/tests/support/tcp.rs b/proxy/tests/support/tcp.rs
index 87256b4b0..bb4388d72 100644
--- a/proxy/tests/support/tcp.rs
+++ b/proxy/tests/support/tcp.rs
@@ -26,8 +26,20 @@ pub struct TcpClient {
     tx: TcpSender,
 }
 
+type Handler = Box<CallBox + Send>;
+
+trait CallBox: 'static {
+    fn call_box(self: Box<Self>, sock: TcpStream) -> Box<Future<Item=(), Error=()>>;
+}
+
+impl<F: FnOnce(TcpStream) -> Box<Future<Item=(), Error=()>> + Send + 'static> CallBox for F {
+    fn call_box(self: Box<Self>, sock: TcpStream) -> Box<Future<Item=(), Error=()>> {
+        (*self)(sock)
+    }
+}
+
 pub struct TcpServer {
-    accepts: VecDeque<Box<Fn(Vec<u8>) -> Vec<u8> + Send>>,
+    accepts: VecDeque<Handler>,
 }
 
 pub struct TcpConn {
@@ -50,10 +62,29 @@ impl TcpClient {
 impl TcpServer {
     pub fn accept<F, U>(mut self, cb: F) -> Self
     where
-        F: Fn(Vec<u8>) -> U + Send + 'static,
+        F: FnOnce(Vec<u8>) -> U + Send + 'static,
         U: Into<Vec<u8>>,
     {
-        self.accepts.push_back(Box::new(move |v| cb(v).into()));
+        self.accept_fut(move |sock| {
+            tokio_io::io::read(sock, vec![0; 1024])
+                .and_then(move |(sock, mut vec, n)| {
+                    vec.truncate(n);
+                    let write = cb(vec).into();
+                    tokio_io::io::write_all(sock, write)
+                })
+                .map(|_| ())
+                .map_err(|e| panic!("tcp server error: {}", e))
+        })
+    }
+
+    pub fn accept_fut<F, U>(mut self, cb: F) -> Self
+    where
+        F: FnOnce(TcpStream) -> U + Send + 'static,
+        U: IntoFuture<Item=(), Error=()> + 'static,
+    {
+        self.accepts.push_back(Box::new(move |tcp| -> Box<Future<Item=(), Error=()>> {
+            Box::new(cb(tcp).into_future())
+        }));
         self
     }
 
@@ -166,15 +197,7 @@ fn run_server(tcp: TcpServer) -> server::Listening {
         let work = bind.incoming().for_each(move |(sock, _)| {
             let cb = accepts.pop_front().expect("no more accepts");
 
-            let fut = tokio_io::io::read(sock, vec![0; 1024])
-                .and_then(move |(sock, mut vec, n)| {
-                    vec.truncate(n);
-                    let write = cb(vec);
-                    tokio_io::io::write_all(sock, write)
-                })
-                .map(|_| ())
-                .map_err(|e| panic!("tcp server error: {}", e));
-
+            let fut = cb.call_box(sock);
             reactor.spawn(fut);
             Ok(())
         });
diff --git a/proxy/tests/transparency.rs b/proxy/tests/transparency.rs
index 17e287142..373a5d335 100644
--- a/proxy/tests/transparency.rs
+++ b/proxy/tests/transparency.rs
@@ -249,6 +249,55 @@ fn tcp_with_no_orig_dst() {
     assert_eq!(read, b"");
 }
 
+#[test]
+fn tcp_connections_close_if_client_closes() {
+    use std::sync::mpsc;
+
+    let _ = env_logger::try_init();
+
+    let msg1 = "custom tcp hello";
+    let msg2 = "custom tcp bye";
+
+    let (tx, rx) = mpsc::channel();
+
+    let srv = server::tcp()
+        .accept_fut(move |sock| {
+            tokio_io::io::read(sock, vec![0; 1024])
+                .and_then(move |(sock, vec, n)| {
+                    assert_eq!(&vec[..n], msg1.as_bytes());
+
+                    tokio_io::io::write_all(sock, msg2.as_bytes())
+                }).and_then(|(sock, _)| {
+                    // lets read again, but we should get eof
+                    tokio_io::io::read(sock, [0; 16])
+                })
+                .map(move |(_sock, _vec, n)| {
+                    assert_eq!(n, 0);
+                    tx.send(()).unwrap();
+                })
+                .map_err(|e| panic!("tcp server error: {}", e))
+        })
+        .run();
+    let ctrl = controller::new().run();
+    let proxy = proxy::new()
+        .controller(ctrl)
+        .inbound(srv)
+        .run();
+
+    let client = client::tcp(proxy.inbound);
+
+    let tcp_client = client.connect();
+    tcp_client.write(msg1);
+    assert_eq!(tcp_client.read(), msg2.as_bytes());
+
+    drop(tcp_client);
+
+    // rx will be fulfilled when our tcp accept_fut sees
+    // a socket disconnect, which is what we are testing for.
+    // the timeout here is just to prevent this test from hanging
+    rx.recv_timeout(Duration::from_secs(5)).unwrap();
+}
+
 #[test]
 fn http11_upgrade_not_supported() {
     let _ = env_logger::try_init();