Rewrite lightning-net-tokio using async/await and tokio 0.2
authorMatt Corallo <git@bluematt.me>
Sat, 1 Feb 2020 17:27:30 +0000 (12:27 -0500)
committerMatt Corallo <git@bluematt.me>
Wed, 11 Mar 2020 16:19:39 +0000 (12:19 -0400)
This is a rather major rewrite, using async/await and tokio 0.2,
which cleans up the code a ton as well as adds significantly to
readability.

lightning-net-tokio/Cargo.toml
lightning-net-tokio/src/lib.rs

index 1bd9805a7c901bdb609e82f43bb6ae8c7d04013b..c915220d64252d878b4c16083347743347b33524 100644 (file)
@@ -1,8 +1,9 @@
 [package]
 name = "lightning-net-tokio"
-version = "0.0.2"
+version = "0.0.3"
 authors = ["Matt Corallo"]
 license = "Apache-2.0"
+edition = "2018"
 description = """
 Implementation of the rust-lightning network stack using Tokio.
 For Rust-Lightning clients which wish to make direct connections to Lightning P2P nodes, this is a simple alternative to implementing the nerequired network stack, especially for those already using Tokio.
@@ -13,7 +14,7 @@ bitcoin = "0.21"
 bitcoin_hashes = "0.7"
 lightning = { version = "0.0.10", path = "../lightning" }
 secp256k1 = "0.15"
-tokio-codec = "0.1"
-futures = "0.1"
-tokio = "0.1"
-bytes = "0.4"
+tokio = { version = ">=0.2.12", features = [ "io-util", "macros", "rt-core", "sync", "tcp", "time" ] }
+
+[dev-dependencies]
+tokio = { version = ">=0.2.12", features = [ "io-util", "macros", "rt-core", "rt-threaded", "sync", "tcp", "time" ] }
index d4621e0003c3d2c5f0e67ef1dc4a118791541166..08aa12571985a985f896929e2034a248ff77ca1a 100644 (file)
-extern crate bytes;
-extern crate tokio;
-extern crate tokio_codec;
-extern crate futures;
-extern crate lightning;
-extern crate secp256k1;
-
-use bytes::BufMut;
-
-use futures::future;
-use futures::future::Future;
-use futures::{AsyncSink, Stream, Sink};
-use futures::sync::mpsc;
+//! A socket handling library for those running in Tokio environments who wish to use
+//! rust-lightning with native TcpStreams.
+//!
+//! Designed to be as simple as possible, the high-level usage is almost as simple as "hand over a
+//! TcpStream and a reference to a PeerManager and the rest is handled", except for the
+//! [Event](../lightning/util/events/enum.Event.html) handlng mechanism, see below.
+//!
+//! The PeerHandler, due to the fire-and-forget nature of this logic, must be an Arc, and must use
+//! the SocketDescriptor provided here as the PeerHandler's SocketDescriptor.
+//!
+//! Three methods are exposed to register a new connection for handling in tokio::spawn calls, see
+//! their individual docs for more. All three take a
+//! [mpsc::Sender<()>](../tokio/sync/mpsc/struct.Sender.html) which is sent into every time
+//! something occurs which may result in lightning [Events](../lightning/util/events/enum.Event.html).
+//! The call site should, thus, look something like this:
+//! ```
+//! use tokio::sync::mpsc;
+//! use tokio::net::TcpStream;
+//! use secp256k1::key::PublicKey;
+//! use lightning::util::events::EventsProvider;
+//! use std::net::SocketAddr;
+//! use std::sync::Arc;
+//!
+//! // Define concrete types for our high-level objects:
+//! type TxBroadcaster = dyn lightning::chain::chaininterface::BroadcasterInterface;
+//! type FeeEstimator = dyn lightning::chain::chaininterface::FeeEstimator;
+//! type ChannelMonitor = lightning::ln::channelmonitor::SimpleManyChannelMonitor<lightning::chain::transaction::OutPoint, lightning::chain::keysinterface::InMemoryChannelKeys, Arc<TxBroadcaster>, Arc<FeeEstimator>>;
+//! type ChannelManager = lightning::ln::channelmanager::SimpleArcChannelManager<ChannelMonitor, TxBroadcaster, FeeEstimator>;
+//! type PeerManager = lightning::ln::peer_handler::SimpleArcPeerManager<lightning_net_tokio::SocketDescriptor, ChannelMonitor, TxBroadcaster, FeeEstimator>;
+//!
+//! // Connect to node with pubkey their_node_id at addr:
+//! async fn connect_to_node(peer_manager: PeerManager, channel_monitor: Arc<ChannelMonitor>, channel_manager: ChannelManager, their_node_id: PublicKey, addr: SocketAddr) {
+//!     let (sender, mut receiver) = mpsc::channel(2);
+//!     lightning_net_tokio::connect_outbound(peer_manager, sender, their_node_id, addr).await;
+//!     loop {
+//!         receiver.recv().await;
+//!         for _event in channel_manager.get_and_clear_pending_events().drain(..) {
+//!             // Handle the event!
+//!         }
+//!         for _event in channel_monitor.get_and_clear_pending_events().drain(..) {
+//!             // Handle the event!
+//!         }
+//!     }
+//! }
+//!
+//! // Begin reading from a newly accepted socket and talk to the peer:
+//! async fn accept_socket(peer_manager: PeerManager, channel_monitor: Arc<ChannelMonitor>, channel_manager: ChannelManager, socket: TcpStream) {
+//!     let (sender, mut receiver) = mpsc::channel(2);
+//!     lightning_net_tokio::setup_inbound(peer_manager, sender, socket);
+//!     loop {
+//!         receiver.recv().await;
+//!         for _event in channel_manager.get_and_clear_pending_events().drain(..) {
+//!             // Handle the event!
+//!         }
+//!         for _event in channel_monitor.get_and_clear_pending_events().drain(..) {
+//!             // Handle the event!
+//!         }
+//!     }
+//! }
+//! ```
 
 use secp256k1::key::PublicKey;
 
-use tokio::timer::Delay;
 use tokio::net::TcpStream;
+use tokio::{io, time};
+use tokio::sync::mpsc;
+use tokio::io::{AsyncReadExt, AsyncWrite, AsyncWriteExt};
 
 use lightning::ln::peer_handler;
 use lightning::ln::peer_handler::SocketDescriptor as LnSocketTrait;
 use lightning::ln::msgs::ChannelMessageHandler;
 
-use std::mem;
+use std::task;
 use std::net::SocketAddr;
-use std::sync::{Arc, Mutex};
+use std::sync::{Arc, Mutex, MutexGuard};
 use std::sync::atomic::{AtomicU64, Ordering};
-use std::time::{Duration, Instant};
-use std::vec::Vec;
+use std::time::Duration;
 use std::hash::Hash;
 
 static ID_COUNTER: AtomicU64 = AtomicU64::new(0);
 
-/// A connection to a remote peer. Can be constructed either as a remote connection using
-/// Connection::setup_outbound o
-pub struct Connection {
-       writer: Option<mpsc::Sender<bytes::Bytes>>,
+/// Connection contains all our internal state for a connection - we hold a reference to the
+/// Connection object (in an Arc<Mutex<>>) in each SocketDescriptor we create as well as in the
+/// read future (which is returned by schedule_read).
+struct Connection {
+       writer: Option<io::WriteHalf<TcpStream>>,
        event_notify: mpsc::Sender<()>,
-       pending_read: Vec<u8>,
-       read_blocker: Option<futures::sync::oneshot::Sender<Result<(), ()>>>,
+       // Because our PeerManager is templated by user-provided types, and we can't (as far as I can
+       // tell) have a const RawWakerVTable built out of templated functions, we need some indirection
+       // between being woken up with write-ready and calling PeerManager::write_buffer_spce_avail.
+       // This provides that indirection, with a Sender which gets handed to the PeerManager Arc on
+       // the schedule_read stack.
+       //
+       // An alternative (likely more effecient) approach would involve creating a RawWakerVTable at
+       // runtime with functions templated by the Arc<PeerManager> type, calling
+       // write_buffer_space_avail directly from tokio's write wake, however doing so would require
+       // more unsafe voodo than I really feel like writing.
+       write_avail: mpsc::Sender<()>,
+       // When we are told by rust-lightning to pause read (because we have writes backing up), we do
+       // so by setting read_paused. At that point, the read task will stop reading bytes from the
+       // socket. To wake it up (without otherwise changing its state, we can push a value into this
+       // Sender.
+       read_waker: mpsc::Sender<()>,
        read_paused: bool,
-       need_disconnect: bool,
+       rl_requested_disconnect: bool,
        id: u64,
 }
 impl Connection {
-       fn schedule_read<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor<CMH>, Arc<CMH>>>, us: Arc<Mutex<Self>>, reader: futures::stream::SplitStream<tokio_codec::Framed<TcpStream, tokio_codec::BytesCodec>>) {
-               let us_ref = us.clone();
-               let us_close_ref = us.clone();
+       fn event_trigger(us: &mut MutexGuard<Self>) {
+               match us.event_notify.try_send(()) {
+                       Ok(_) => {},
+                       Err(mpsc::error::TrySendError::Full(_)) => {
+                               // Ignore full errors as we just need the user to poll after this point, so if they
+                               // haven't received the last send yet, it doesn't matter.
+                       },
+                       _ => panic!()
+               }
+       }
+       async fn schedule_read<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor, Arc<CMH>>>, us: Arc<Mutex<Self>>, mut reader: io::ReadHalf<TcpStream>, mut read_wake_receiver: mpsc::Receiver<()>, mut write_avail_receiver: mpsc::Receiver<()>) {
                let peer_manager_ref = peer_manager.clone();
-               tokio::spawn(reader.for_each(move |b| {
-                       let pending_read = b.to_vec();
-                       {
-                               let mut lock = us_ref.lock().unwrap();
-                               assert!(lock.pending_read.is_empty());
-                               if lock.read_paused {
-                                       lock.pending_read = pending_read;
-                                       let (sender, blocker) = futures::sync::oneshot::channel();
-                                       lock.read_blocker = Some(sender);
-                                       return future::Either::A(blocker.then(|_| { Ok(()) }));
-                               }
+               // 8KB is nice and big but also should never cause any issues with stack overflowing.
+               let mut buf = [0; 8192];
+
+               let mut our_descriptor = SocketDescriptor::new(us.clone());
+               // An enum describing why we did/are disconnecting:
+               enum Disconnect {
+                       // Rust-Lightning told us to disconnect, either by returning an Err or by calling
+                       // SocketDescriptor::disconnect_socket.
+                       // In this case, we do not call peer_manager.socket_disconnected() as Rust-Lightning
+                       // already knows we're disconnected.
+                       CloseConnection,
+                       // The connection was disconnected for some other reason, ie because the socket was
+                       // closed.
+                       // In this case, we do need to call peer_manager.socket_disconnected() to inform
+                       // Rust-Lightning that the socket is gone.
+                       PeerDisconnected
+               };
+               let disconnect_type = loop {
+                       macro_rules! shutdown_socket {
+                               ($err: expr, $need_disconnect: expr) => { {
+                                       println!("Disconnecting peer due to {}!", $err);
+                                       break $need_disconnect;
+                               } }
                        }
-                       //TODO: There's a race where we don't meet the requirements of socket_disconnected if its
-                       //called right here, after we release the us_ref lock in the scope above, but before we
-                       //call read_event!
-                       match peer_manager.read_event(&mut SocketDescriptor::new(us_ref.clone(), peer_manager.clone()), &pending_read) {
-                               Ok(pause_read) => {
-                                       if pause_read {
-                                               let mut lock = us_ref.lock().unwrap();
-                                               lock.read_paused = true;
+
+                       let read_paused = us.lock().unwrap().read_paused;
+                       tokio::select! {
+                               v = write_avail_receiver.recv() => {
+                                       assert!(v.is_some()); // We can't have dropped the sending end, its in the us Arc!
+                                       if us.lock().unwrap().rl_requested_disconnect {
+                                               shutdown_socket!("disconnect_socket() call from RL", Disconnect::CloseConnection);
+                                       }
+                                       if let Err(e) = peer_manager.write_buffer_space_avail(&mut our_descriptor) {
+                                               shutdown_socket!(e, Disconnect::CloseConnection);
                                        }
                                },
-                               Err(e) => {
-                                       us_ref.lock().unwrap().need_disconnect = false;
-                                       return future::Either::B(future::result(Err(std::io::Error::new(std::io::ErrorKind::InvalidData, e))));
-                               }
+                               _ = read_wake_receiver.recv() => {},
+                               read = reader.read(&mut buf), if !read_paused => match read {
+                                       Ok(0) => shutdown_socket!("Connection closed", Disconnect::PeerDisconnected),
+                                       Ok(len) => {
+                                               if us.lock().unwrap().rl_requested_disconnect {
+                                                       shutdown_socket!("disconnect_socket() call from RL", Disconnect::CloseConnection);
+                                               }
+                                               let read_res = peer_manager.read_event(&mut our_descriptor, &buf[0..len]);
+                                               match read_res {
+                                                       Ok(pause_read) => {
+                                                               let mut us_lock = us.lock().unwrap();
+                                                               if pause_read {
+                                                                       us_lock.read_paused = true;
+                                                               }
+                                                               Self::event_trigger(&mut us_lock);
+                                                       },
+                                                       Err(e) => shutdown_socket!(e, Disconnect::CloseConnection),
+                                               }
+                                       },
+                                       Err(e) => shutdown_socket!(e, Disconnect::PeerDisconnected),
+                               },
                        }
+               };
+               let writer_option = us.lock().unwrap().writer.take();
+               if let Some(mut writer) = writer_option {
+                       // If the socket is already closed, shutdown() will fail, so just ignore it.
+                       let _ = writer.shutdown().await;
+               }
+               if let Disconnect::PeerDisconnected = disconnect_type {
+                       peer_manager_ref.socket_disconnected(&our_descriptor);
+                       Self::event_trigger(&mut us.lock().unwrap());
+               }
+       }
 
-                       if let Err(e) = us_ref.lock().unwrap().event_notify.try_send(()) {
-                               // Ignore full errors as we just need them to poll after this point, so if the user
-                               // hasn't received the last send yet, it doesn't matter.
-                               assert!(e.is_full());
-                       }
+       fn new(event_notify: mpsc::Sender<()>, stream: TcpStream) -> (io::ReadHalf<TcpStream>, mpsc::Receiver<()>, mpsc::Receiver<()>, Arc<Mutex<Self>>) {
+               // We only ever need a channel of depth 1 here: if we returned a non-full write to the
+               // PeerManager, we will eventually get notified that there is room in the socket to write
+               // new bytes, which will generate an event. That event will be popped off the queue before
+               // we call write_buffer_space_avail, ensuring that we have room to push a new () if, during
+               // the write_buffer_space_avail() call, send_data() returns a non-full write.
+               let (write_avail, write_receiver) = mpsc::channel(1);
+               // Similarly here - our only goal is to make sure the reader wakes up at some point after
+               // we shove a value into the channel which comes after we've reset the read_paused bool to
+               // false.
+               let (read_waker, read_receiver) = mpsc::channel(1);
+               let (reader, writer) = io::split(stream);
 
-                       future::Either::B(future::result(Ok(())))
-               }).then(move |_| {
-                       if us_close_ref.lock().unwrap().need_disconnect {
-                               peer_manager_ref.socket_disconnected(&SocketDescriptor::new(us_close_ref, peer_manager_ref.clone()));
-                               println!("Peer disconnected!");
-                       } else {
-                               println!("We disconnected peer!");
-                       }
-                       Ok(())
-               }));
+               (reader, write_receiver, read_receiver,
+               Arc::new(Mutex::new(Self {
+                       writer: Some(writer), event_notify, write_avail, read_waker,
+                       read_paused: false, rl_requested_disconnect: false,
+                       id: ID_COUNTER.fetch_add(1, Ordering::AcqRel)
+               })))
        }
+}
 
-       fn new(event_notify: mpsc::Sender<()>, stream: TcpStream) -> (futures::stream::SplitStream<tokio_codec::Framed<TcpStream, tokio_codec::BytesCodec>>, Arc<Mutex<Self>>) {
-               let (writer, reader) = tokio_codec::Framed::new(stream, tokio_codec::BytesCodec::new()).split();
-               let (send_sink, send_stream) = mpsc::channel(3);
-               tokio::spawn(writer.send_all(send_stream.map_err(|_| -> std::io::Error {
-                       unreachable!();
-               })).then(|_| {
-                       future::result(Ok(()))
-               }));
-               let us = Arc::new(Mutex::new(Self { writer: Some(send_sink), event_notify, pending_read: Vec::new(), read_blocker: None, read_paused: false, need_disconnect: true, id: ID_COUNTER.fetch_add(1, Ordering::AcqRel) }));
-
-               (reader, us)
-       }
+/// Process incoming messages and feed outgoing messages on the provided socket generated by
+/// accepting an incoming connection.
+///
+/// The returned future will complete when the peer is disconnected and associated handling
+/// futures are freed, though, because all processing futures are spawned with tokio::spawn, you do
+/// not need to poll the provided future in order to make progress.
+///
+/// See the module-level documentation for how to handle the event_notify mpsc::Sender.
+pub fn setup_inbound<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor, Arc<CMH>>>, event_notify: mpsc::Sender<()>, stream: TcpStream) -> impl std::future::Future<Output=()> {
+       let (reader, write_receiver, read_receiver, us) = Connection::new(event_notify, stream);
+       #[cfg(debug_assertions)]
+       let last_us = Arc::clone(&us);
 
-       /// Process incoming messages and feed outgoing messages on the provided socket generated by
-       /// accepting an incoming connection (by scheduling futures with tokio::spawn).
-       ///
-       /// You should poll the Receive end of event_notify and call get_and_clear_pending_events() on
-       /// ChannelManager and ChannelMonitor objects.
-       pub fn setup_inbound<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor<CMH>, Arc<CMH>>>, event_notify: mpsc::Sender<()>, stream: TcpStream) {
-               let (reader, us) = Self::new(event_notify, stream);
+       let handle_opt = if let Ok(_) = peer_manager.new_inbound_connection(SocketDescriptor::new(us.clone())) {
+               Some(tokio::spawn(Connection::schedule_read(peer_manager, us, reader, read_receiver, write_receiver)))
+       } else {
+               // Note that we will skip socket_disconnected here, in accordance with the PeerManager
+               // requirements.
+               None
+       };
 
-               if let Ok(_) = peer_manager.new_inbound_connection(SocketDescriptor::new(us.clone(), peer_manager.clone())) {
-                       Self::schedule_read(peer_manager, us, reader);
+       async move {
+               if let Some(handle) = handle_opt {
+                       if let Err(e) = handle.await {
+                               assert!(e.is_cancelled());
+                       } else {
+                               // This is certainly not guaranteed to always be true - the read loop may exit
+                               // while there are still pending write wakers that need to be woken up after the
+                               // socket shutdown(). Still, as a check during testing, to make sure tokio doesn't
+                               // keep too many wakers around, this makes sense. The race should be rare (we do
+                               // some work after shutdown()) and an error would be a major memory leak.
+                               #[cfg(debug_assertions)]
+                               assert!(Arc::try_unwrap(last_us).is_ok());
+                       }
                }
        }
+}
+
+/// Process incoming messages and feed outgoing messages on the provided socket generated by
+/// making an outbound connection which is expected to be accepted by a peer with the given
+/// public key. The relevant processing is set to run free (via tokio::spawn).
+///
+/// The returned future will complete when the peer is disconnected and associated handling
+/// futures are freed, though, because all processing futures are spawned with tokio::spawn, you do
+/// not need to poll the provided future in order to make progress.
+///
+/// See the module-level documentation for how to handle the event_notify mpsc::Sender.
+pub fn setup_outbound<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor, Arc<CMH>>>, event_notify: mpsc::Sender<()>, their_node_id: PublicKey, stream: TcpStream) -> impl std::future::Future<Output=()> {
+       let (reader, write_receiver, read_receiver, us) = Connection::new(event_notify, stream);
+       #[cfg(debug_assertions)]
+       let last_us = Arc::clone(&us);
+
+       let handle_opt = if let Ok(initial_send) = peer_manager.new_outbound_connection(their_node_id, SocketDescriptor::new(us.clone())) {
+               Some(tokio::spawn(async move {
+                       if SocketDescriptor::new(us.clone()).send_data(&initial_send, true) != initial_send.len() {
+                               // We should essentially always have enough room in a TCP socket buffer to send the
+                               // initial 10s of bytes, if not, just give up as hopeless.
+                               eprintln!("Failed to write first full message to socket!");
+                               peer_manager.socket_disconnected(&SocketDescriptor::new(Arc::clone(&us)));
+                       } else {
+                               Connection::schedule_read(peer_manager, us, reader, read_receiver, write_receiver).await;
+                       }
+               }))
+       } else {
+               // Note that we will skip socket_disconnected here, in accordance with the PeerManager
+               // requirements.
+               None
+       };
 
-       /// Process incoming messages and feed outgoing messages on the provided socket generated by
-       /// making an outbound connection which is expected to be accepted by a peer with the given
-       /// public key (by scheduling futures with tokio::spawn).
-       ///
-       /// You should poll the Receive end of event_notify and call get_and_clear_pending_events() on
-       /// ChannelManager and ChannelMonitor objects.
-       pub fn setup_outbound<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor<CMH>, Arc<CMH>>>, event_notify: mpsc::Sender<()>, their_node_id: PublicKey, stream: TcpStream) {
-               let (reader, us) = Self::new(event_notify, stream);
-
-               if let Ok(initial_send) = peer_manager.new_outbound_connection(their_node_id, SocketDescriptor::new(us.clone(), peer_manager.clone())) {
-                       if SocketDescriptor::new(us.clone(), peer_manager.clone()).send_data(&initial_send, true) == initial_send.len() {
-                               Self::schedule_read(peer_manager, us, reader);
+       async move {
+               if let Some(handle) = handle_opt {
+                       if let Err(e) = handle.await {
+                               assert!(e.is_cancelled());
                        } else {
-                               println!("Failed to write first full message to socket!");
+                               // This is certainly not guaranteed to always be true - the read loop may exit
+                               // while there are still pending write wakers that need to be woken up after the
+                               // socket shutdown(). Still, as a check during testing, to make sure tokio doesn't
+                               // keep too many wakers around, this makes sense. The race should be rare (we do
+                               // some work after shutdown()) and an error would be a major memory leak.
+                               #[cfg(debug_assertions)]
+                               assert!(Arc::try_unwrap(last_us).is_ok());
                        }
                }
        }
+}
 
-       /// Process incoming messages and feed outgoing messages on a new connection made to the given
-       /// socket address which is expected to be accepted by a peer with the given public key (by
-       /// scheduling futures with tokio::spawn).
-       ///
-       /// You should poll the Receive end of event_notify and call get_and_clear_pending_events() on
-       /// ChannelManager and ChannelMonitor objects.
-       pub fn connect_outbound<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor<CMH>, Arc<CMH>>>, event_notify: mpsc::Sender<()>, their_node_id: PublicKey, addr: SocketAddr) {
-               let connect_timeout = Delay::new(Instant::now() + Duration::from_secs(10)).then(|_| {
-                       future::err(std::io::Error::new(std::io::ErrorKind::TimedOut, "timeout reached"))
-               });
-               tokio::spawn(TcpStream::connect(&addr).select(connect_timeout)
-                       .and_then(move |stream| {
-                               Connection::setup_outbound(peer_manager, event_notify, their_node_id, stream.0);
-                               future::ok(())
-                       }).or_else(|_| {
-                               //TODO: return errors somehow
-                               future::ok(())
-                       }));
-       }
+/// Process incoming messages and feed outgoing messages on a new connection made to the given
+/// socket address which is expected to be accepted by a peer with the given public key (by
+/// scheduling futures with tokio::spawn).
+///
+/// Shorthand for TcpStream::connect(addr) with a timeout followed by setup_outbound().
+///
+/// Returns a future (as the fn is async) which needs to be polled to complete the connection and
+/// connection setup. That future then returns a future which will complete when the peer is
+/// disconnected and associated handling futures are freed, though, because all processing in said
+/// futures are spawned with tokio::spawn, you do not need to poll the second future in order to
+/// make progress.
+///
+/// See the module-level documentation for how to handle the event_notify mpsc::Sender.
+pub async fn connect_outbound<CMH: ChannelMessageHandler + 'static>(peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor, Arc<CMH>>>, event_notify: mpsc::Sender<()>, their_node_id: PublicKey, addr: SocketAddr) -> Option<impl std::future::Future<Output=()>> {
+       if let Ok(Ok(stream)) = time::timeout(Duration::from_secs(10), TcpStream::connect(&addr)).await {
+               Some(setup_outbound(peer_manager, event_notify, their_node_id, stream))
+       } else { None }
 }
 
-pub struct SocketDescriptor<CMH: ChannelMessageHandler + 'static> {
+const SOCK_WAKER_VTABLE: task::RawWakerVTable =
+       task::RawWakerVTable::new(clone_socket_waker, wake_socket_waker, wake_socket_waker_by_ref, drop_socket_waker);
+
+fn clone_socket_waker(orig_ptr: *const ()) -> task::RawWaker {
+       write_avail_to_waker(orig_ptr as *const mpsc::Sender<()>)
+}
+// When waking, an error should be fine. Most likely we got two send_datas in a row, both of which
+// failed to fully write, but we only need to call write_buffer_space_avail() once. Otherwise, the
+// sending thread may have already gone away due to a socket close, in which case there's nothing
+// to wake up anyway.
+fn wake_socket_waker(orig_ptr: *const ()) {
+       let sender = unsafe { &mut *(orig_ptr as *mut mpsc::Sender<()>) };
+       let _ = sender.try_send(());
+       drop_socket_waker(orig_ptr);
+}
+fn wake_socket_waker_by_ref(orig_ptr: *const ()) {
+       let sender_ptr = orig_ptr as *const mpsc::Sender<()>;
+       let mut sender = unsafe { (*sender_ptr).clone() };
+       let _ = sender.try_send(());
+}
+fn drop_socket_waker(orig_ptr: *const ()) {
+       let _orig_box = unsafe { Box::from_raw(orig_ptr as *mut mpsc::Sender<()>) };
+       // _orig_box is now dropped
+}
+fn write_avail_to_waker(sender: *const mpsc::Sender<()>) -> task::RawWaker {
+       let new_box = Box::leak(Box::new(unsafe { (*sender).clone() }));
+       let new_ptr = new_box as *const mpsc::Sender<()>;
+       task::RawWaker::new(new_ptr as *const (), &SOCK_WAKER_VTABLE)
+}
+
+/// The SocketDescriptor used to refer to sockets by a PeerHandler. This is pub only as it is a
+/// type in the template of PeerHandler.
+pub struct SocketDescriptor {
        conn: Arc<Mutex<Connection>>,
        id: u64,
-       peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor<CMH>, Arc<CMH>>>,
 }
-impl<CMH: ChannelMessageHandler> SocketDescriptor<CMH> {
-       fn new(conn: Arc<Mutex<Connection>>, peer_manager: Arc<peer_handler::PeerManager<SocketDescriptor<CMH>, Arc<CMH>>>) -> Self {
+impl SocketDescriptor {
+       fn new(conn: Arc<Mutex<Connection>>) -> Self {
                let id = conn.lock().unwrap().id;
-               Self { conn, id, peer_manager }
+               Self { conn, id }
        }
 }
-impl<CMH: ChannelMessageHandler> peer_handler::SocketDescriptor for SocketDescriptor<CMH> {
+impl peer_handler::SocketDescriptor for SocketDescriptor {
        fn send_data(&mut self, data: &[u8], resume_read: bool) -> usize {
-               macro_rules! schedule_read {
-                       ($us_ref: expr) => {
-                               tokio::spawn(future::lazy(move || -> Result<(), ()> {
-                                       let mut read_data = Vec::new();
-                                       {
-                                               let mut us = $us_ref.conn.lock().unwrap();
-                                               mem::swap(&mut read_data, &mut us.pending_read);
-                                       }
-                                       if !read_data.is_empty() {
-                                               let mut us_clone = $us_ref.clone();
-                                               match $us_ref.peer_manager.read_event(&mut us_clone, &read_data) {
-                                                       Ok(pause_read) => {
-                                                               if pause_read { return Ok(()); }
-                                                       },
-                                                       Err(_) => {
-                                                               //TODO: Not actually sure how to do this
-                                                               return Ok(());
-                                                       }
-                                               }
-                                       }
-                                       let mut us = $us_ref.conn.lock().unwrap();
-                                       if let Some(sender) = us.read_blocker.take() {
-                                               sender.send(Ok(())).unwrap();
-                                       }
-                                       us.read_paused = false;
-                                       if let Err(e) = us.event_notify.try_send(()) {
-                                               // Ignore full errors as we just need them to poll after this point, so if the user
-                                               // hasn't received the last send yet, it doesn't matter.
-                                               assert!(e.is_full());
-                                       }
-                                       Ok(())
-                               }));
-                       }
-               }
-
+               // To send data, we take a lock on our Connection to access the WriteHalf of the TcpStream,
+               // writing to it if there's room in the kernel buffer, or otherwise create a new Waker with
+               // a SocketDescriptor in it which can wake up the write_avail Sender, waking up the
+               // processing future which will call write_buffer_space_avail and we'll end up back here.
                let mut us = self.conn.lock().unwrap();
-               if resume_read {
-                       let us_ref = self.clone();
-                       schedule_read!(us_ref);
-               }
-               if data.is_empty() { return 0; }
                if us.writer.is_none() {
-                       us.read_paused = true;
+                       // The writer gets take()n when it is time to shut down, so just fast-return 0 here.
                        return 0;
                }
 
-               let mut bytes = bytes::BytesMut::with_capacity(data.len());
-               bytes.put(data);
-               let write_res = us.writer.as_mut().unwrap().start_send(bytes.freeze());
-               match write_res {
-                       Ok(res) => {
-                               match res {
-                                       AsyncSink::Ready => {
-                                               data.len()
-                                       },
-                                       AsyncSink::NotReady(_) => {
-                                               us.read_paused = true;
-                                               let us_ref = self.clone();
-                                               tokio::spawn(us.writer.take().unwrap().flush().then(move |writer_res| -> Result<(), ()> {
-                                                       if let Ok(writer) = writer_res {
-                                                               {
-                                                                       let mut us = us_ref.conn.lock().unwrap();
-                                                                       us.writer = Some(writer);
-                                                               }
-                                                               schedule_read!(us_ref);
-                                                       } // we'll fire the disconnect event on the socket reader end
-                                                       Ok(())
-                                               }));
-                                               0
-                                       }
-                               }
-                       },
-                       Err(_) => {
-                               // We'll fire the disconnected event on the socket reader end
-                               0
-                       },
+               if resume_read && us.read_paused {
+                       // The schedule_read future may go to lock up but end up getting woken up by there
+                       // being more room in the write buffer, dropping the other end of this Sender
+                       // before we get here, so we ignore any failures to wake it up.
+                       us.read_paused = false;
+                       let _ = us.read_waker.try_send(());
+               }
+               if data.is_empty() { return 0; }
+               let waker = unsafe { task::Waker::from_raw(write_avail_to_waker(&us.write_avail)) };
+               let mut ctx = task::Context::from_waker(&waker);
+               let mut written_len = 0;
+               loop {
+                       match std::pin::Pin::new(us.writer.as_mut().unwrap()).poll_write(&mut ctx, &data[written_len..]) {
+                               task::Poll::Ready(Ok(res)) => {
+                                       // The tokio docs *seem* to indicate this can't happen, and I certainly don't
+                                       // know how to handle it if it does (cause it should be a Poll::Pending
+                                       // instead):
+                                       assert_ne!(res, 0);
+                                       written_len += res;
+                                       if written_len == data.len() { return written_len; }
+                               },
+                               task::Poll::Ready(Err(e)) => {
+                                       // The tokio docs *seem* to indicate this can't happen, and I certainly don't
+                                       // know how to handle it if it does (cause it should be a Poll::Pending
+                                       // instead):
+                                       assert_ne!(e.kind(), io::ErrorKind::WouldBlock);
+                                       // Probably we've already been closed, just return what we have and let the
+                                       // read thread handle closing logic.
+                                       return written_len;
+                               },
+                               task::Poll::Pending => {
+                                       // We're queued up for a write event now, but we need to make sure we also
+                                       // pause read given we're now waiting on the remote end to ACK (and in
+                                       // accordance with the send_data() docs).
+                                       us.read_paused = true;
+                                       return written_len;
+                               },
+                       }
                }
        }
 
        fn disconnect_socket(&mut self) {
                let mut us = self.conn.lock().unwrap();
-               us.need_disconnect = true;
+               us.rl_requested_disconnect = true;
                us.read_paused = true;
+               // Wake up the sending thread, assuming it is still alive
+               let _ = us.write_avail.try_send(());
+               // TODO: There's a race where we don't meet the requirements of disconnect_socket if the
+               // read task is about to call a PeerManager function (eg read_event or write_event).
+               // Ideally we need to release the us lock and block until we have confirmation from the
+               // read task that it has broken out of its main loop.
        }
 }
-impl<CMH: ChannelMessageHandler> Clone for SocketDescriptor<CMH> {
+impl Clone for SocketDescriptor {
        fn clone(&self) -> Self {
                Self {
                        conn: Arc::clone(&self.conn),
                        id: self.id,
-                       peer_manager: Arc::clone(&self.peer_manager),
                }
        }
 }
-impl<CMH: ChannelMessageHandler> Eq for SocketDescriptor<CMH> {}
-impl<CMH: ChannelMessageHandler> PartialEq for SocketDescriptor<CMH> {
+impl Eq for SocketDescriptor {}
+impl PartialEq for SocketDescriptor {
        fn eq(&self, o: &Self) -> bool {
                self.id == o.id
        }
 }
-impl<CMH: ChannelMessageHandler> Hash for SocketDescriptor<CMH> {
+impl Hash for SocketDescriptor {
        fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
                self.id.hash(state);
        }
 }
 
+#[cfg(test)]
+mod tests {
+       use lightning::ln::features::*;
+       use lightning::ln::msgs::*;
+       use lightning::ln::peer_handler::{MessageHandler, PeerManager};
+       use lightning::util::events::*;
+       use secp256k1::{Secp256k1, SecretKey, PublicKey};
+
+       use tokio::sync::mpsc;
+
+       use std::mem;
+       use std::sync::{Arc, Mutex};
+       use std::time::Duration;
+
+       pub struct TestLogger();
+       impl lightning::util::logger::Logger for TestLogger {
+               fn log(&self, record: &lightning::util::logger::Record) {
+                       println!("{:<5} [{} : {}, {}] {}", record.level.to_string(), record.module_path, record.file, record.line, record.args);
+               }
+       }
+
+       struct MsgHandler{
+               expected_pubkey: PublicKey,
+               pubkey_connected: mpsc::Sender<()>,
+               pubkey_disconnected: mpsc::Sender<()>,
+               msg_events: Mutex<Vec<MessageSendEvent>>,
+       }
+       impl RoutingMessageHandler for MsgHandler {
+               fn handle_node_announcement(&self, _msg: &NodeAnnouncement) -> Result<bool, LightningError> { Ok(false) }
+               fn handle_channel_announcement(&self, _msg: &ChannelAnnouncement) -> Result<bool, LightningError> { Ok(false) }
+               fn handle_channel_update(&self, _msg: &ChannelUpdate) -> Result<bool, LightningError> { Ok(false) }
+               fn handle_htlc_fail_channel_update(&self, _update: &HTLCFailChannelUpdate) { }
+               fn get_next_channel_announcements(&self, _starting_point: u64, _batch_amount: u8) -> Vec<(ChannelAnnouncement, ChannelUpdate, ChannelUpdate)> { Vec::new() }
+               fn get_next_node_announcements(&self, _starting_point: Option<&PublicKey>, _batch_amount: u8) -> Vec<NodeAnnouncement> { Vec::new() }
+               fn should_request_full_sync(&self, _node_id: &PublicKey) -> bool { false }
+       }
+       impl ChannelMessageHandler for MsgHandler {
+               fn handle_open_channel(&self, _their_node_id: &PublicKey, _their_features: InitFeatures, _msg: &OpenChannel) {}
+               fn handle_accept_channel(&self, _their_node_id: &PublicKey, _their_features: InitFeatures, _msg: &AcceptChannel) {}
+               fn handle_funding_created(&self, _their_node_id: &PublicKey, _msg: &FundingCreated) {}
+               fn handle_funding_signed(&self, _their_node_id: &PublicKey, _msg: &FundingSigned) {}
+               fn handle_funding_locked(&self, _their_node_id: &PublicKey, _msg: &FundingLocked) {}
+               fn handle_shutdown(&self, _their_node_id: &PublicKey, _msg: &Shutdown) {}
+               fn handle_closing_signed(&self, _their_node_id: &PublicKey, _msg: &ClosingSigned) {}
+               fn handle_update_add_htlc(&self, _their_node_id: &PublicKey, _msg: &UpdateAddHTLC) {}
+               fn handle_update_fulfill_htlc(&self, _their_node_id: &PublicKey, _msg: &UpdateFulfillHTLC) {}
+               fn handle_update_fail_htlc(&self, _their_node_id: &PublicKey, _msg: &UpdateFailHTLC) {}
+               fn handle_update_fail_malformed_htlc(&self, _their_node_id: &PublicKey, _msg: &UpdateFailMalformedHTLC) {}
+               fn handle_commitment_signed(&self, _their_node_id: &PublicKey, _msg: &CommitmentSigned) {}
+               fn handle_revoke_and_ack(&self, _their_node_id: &PublicKey, _msg: &RevokeAndACK) {}
+               fn handle_update_fee(&self, _their_node_id: &PublicKey, _msg: &UpdateFee) {}
+               fn handle_announcement_signatures(&self, _their_node_id: &PublicKey, _msg: &AnnouncementSignatures) {}
+               fn peer_disconnected(&self, their_node_id: &PublicKey, _no_connection_possible: bool) {
+                       if *their_node_id == self.expected_pubkey {
+                               self.pubkey_disconnected.clone().try_send(()).unwrap();
+                       }
+               }
+               fn peer_connected(&self, their_node_id: &PublicKey, _msg: &Init) {
+                       if *their_node_id == self.expected_pubkey {
+                               self.pubkey_connected.clone().try_send(()).unwrap();
+                       }
+               }
+               fn handle_channel_reestablish(&self, _their_node_id: &PublicKey, _msg: &ChannelReestablish) {}
+               fn handle_error(&self, _their_node_id: &PublicKey, _msg: &ErrorMessage) {}
+       }
+       impl MessageSendEventsProvider for MsgHandler {
+               fn get_and_clear_pending_msg_events(&self) -> Vec<MessageSendEvent> {
+                       let mut ret = Vec::new();
+                       mem::swap(&mut *self.msg_events.lock().unwrap(), &mut ret);
+                       ret
+               }
+       }
+
+       #[tokio::test(threaded_scheduler)]
+       async fn basic_connection_test() {
+               let secp_ctx = Secp256k1::new();
+               let a_key = SecretKey::from_slice(&[1; 32]).unwrap();
+               let b_key = SecretKey::from_slice(&[1; 32]).unwrap();
+               let a_pub = PublicKey::from_secret_key(&secp_ctx, &a_key);
+               let b_pub = PublicKey::from_secret_key(&secp_ctx, &b_key);
+
+               let (a_connected_sender, mut a_connected) = mpsc::channel(1);
+               let (a_disconnected_sender, mut a_disconnected) = mpsc::channel(1);
+               let a_handler = Arc::new(MsgHandler {
+                       expected_pubkey: b_pub,
+                       pubkey_connected: a_connected_sender,
+                       pubkey_disconnected: a_disconnected_sender,
+                       msg_events: Mutex::new(Vec::new()),
+               });
+               let a_manager = Arc::new(PeerManager::new(MessageHandler {
+                       chan_handler: Arc::clone(&a_handler),
+                       route_handler: Arc::clone(&a_handler) as Arc<dyn RoutingMessageHandler>,
+               }, a_key.clone(), &[1; 32], Arc::new(TestLogger())));
+
+               let (b_connected_sender, mut b_connected) = mpsc::channel(1);
+               let (b_disconnected_sender, mut b_disconnected) = mpsc::channel(1);
+               let b_handler = Arc::new(MsgHandler {
+                       expected_pubkey: a_pub,
+                       pubkey_connected: b_connected_sender,
+                       pubkey_disconnected: b_disconnected_sender,
+                       msg_events: Mutex::new(Vec::new()),
+               });
+               let b_manager = Arc::new(PeerManager::new(MessageHandler {
+                       chan_handler: Arc::clone(&b_handler),
+                       route_handler: Arc::clone(&b_handler) as Arc<dyn RoutingMessageHandler>,
+               }, b_key.clone(), &[2; 32], Arc::new(TestLogger())));
+
+               // We bind on localhost, hoping the environment is properly configured with a local
+               // address. This may not always be the case in containers and the like, so if this test is
+               // failing for you check that you have a loopback interface and it is configured with
+               // 127.0.0.1.
+               let (conn_a, conn_b) = if let Ok(listener) = std::net::TcpListener::bind("127.0.0.1:9735") {
+                       (std::net::TcpStream::connect("127.0.0.1:9735").unwrap(), listener.accept().unwrap().0)
+               } else if let Ok(listener) = std::net::TcpListener::bind("127.0.0.1:9999") {
+                       (std::net::TcpStream::connect("127.0.0.1:9999").unwrap(), listener.accept().unwrap().0)
+               } else if let Ok(listener) = std::net::TcpListener::bind("127.0.0.1:46926") {
+                       (std::net::TcpStream::connect("127.0.0.1:46926").unwrap(), listener.accept().unwrap().0)
+               } else { panic!("Failed to bind to v4 localhost on common ports"); };
+
+               let (sender, _receiver) = mpsc::channel(2);
+               let fut_a = super::setup_outbound(Arc::clone(&a_manager), sender.clone(), b_pub, tokio::net::TcpStream::from_std(conn_a).unwrap());
+               let fut_b = super::setup_inbound(b_manager, sender, tokio::net::TcpStream::from_std(conn_b).unwrap());
+
+               tokio::time::timeout(Duration::from_secs(10), a_connected.recv()).await.unwrap();
+               tokio::time::timeout(Duration::from_secs(1), b_connected.recv()).await.unwrap();
+
+               a_handler.msg_events.lock().unwrap().push(MessageSendEvent::HandleError {
+                       node_id: b_pub, action: ErrorAction::DisconnectPeer { msg: None }
+               });
+               assert!(a_disconnected.try_recv().is_err());
+               assert!(b_disconnected.try_recv().is_err());
+
+               a_manager.process_events();
+               tokio::time::timeout(Duration::from_secs(10), a_disconnected.recv()).await.unwrap();
+               tokio::time::timeout(Duration::from_secs(1), b_disconnected.recv()).await.unwrap();
+
+               fut_a.await;
+               fut_b.await;
+       }
+}