bump to v0.5.4

Reviewed-on: lx/netapp#3

.drone.yml

@@ -1,41 +1,54 @@
+---
 kind: pipeline
 name: default
 
-workspace:
-  base: /drone
-
-clone:
-  disable: true
-
 steps:
-  - name: clone
-    image: alpine/git
-    commands:
-      - mkdir -p cargo
-      - git clone $DRONE_GIT_HTTP_URL
-      - cd netapp
-      - git checkout $DRONE_COMMIT
-
   - name: style
-    image: rustlang/rust:nightly
+    image: rust:1.58-buster
     environment:
       CARGO_HOME: /drone/cargo
+    volumes:
+      - name: cargo
+        path: /drone/cargo
     commands:
-      - cd netapp
+      - rustup component add rustfmt clippy
       - cargo fmt -- --check
       - cargo clippy --all-features -- --deny warnings
       - cargo clippy --example fullmesh -- --deny warnings
       - cargo clippy --example basalt --all-features -- --deny warnings
 
   - name: build
-    image: rustlang/rust:nightly
+    image: rust:1.58-buster
     environment:
       CARGO_HOME: /drone/cargo
+    volumes:
+      - name: cargo
+        path: /drone/cargo
     commands:
       - apt-get update
       - apt-get install --yes libsodium-dev
       - cargo install -f cargo-all-features
-      - cd netapp
       - cargo build-all-features
       - cargo build --example fullmesh
       - cargo build --example basalt --features "basalt"
+
+  - name: test
+    image: rust:1.58-buster
+    environment:
+      CARGO_HOME: /drone/cargo
+    volumes:
+      - name: cargo
+        path: /drone/cargo
+    commands:
+      - apt-get update
+      - apt-get install --yes libsodium-dev
+      - cargo test --all-features -- --test-threads 1
+
+volumes:
+  - name: cargo
+    temp: {}
+---
+kind: signature
+hmac: f0d1a9e8d85a22c1d9084b4d90c9930be9700da52284f1875ece996cc52a6ce9
+
+...

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "netapp"
-version = "0.2.0"
+version = "0.5.4"
 authors = ["Alex Auvolat <alex@adnab.me>"]
 edition = "2018"
 license-file = "LICENSE"
@@ -16,32 +16,38 @@ name = "netapp"
 
 [features]
 default = []
-basalt = ["lru", "rand"]
+basalt = ["lru"]
+telemetry = ["opentelemetry", "opentelemetry-contrib"]
 
 [dependencies]
 futures = "0.3.17"
-tokio = { version = "1.0", default-features = false, features = ["net", "rt", "rt-multi-thread", "sync", "time", "macros", "io-util"] }
-tokio-util = { version = "0.6.8", default-features = false, features = ["compat"] }
+pin-project = "1.0.10"
+tokio = { version = "1.0", default-features = false, features = ["net", "rt", "rt-multi-thread", "sync", "time", "macros", "io-util", "signal"] }
+tokio-util = { version = "0.7", default-features = false, features = ["compat", "io"] }
+tokio-stream = "0.1.7"
 
-serde = { version = "1.0", default-features = false, features = ["derive"] }
-rmp-serde = "0.14.3"
+serde = { version = "1.0", default-features = false, features = ["derive", "rc"] }
+rmp-serde = "0.15"
 hex = "0.4.2"
-base64 = "0.12.1"
 
-rand = { version = "0.5.5", optional = true }
+rand = { version = "0.8" }
 
 log = "0.4.8"
-env_logger = "0.8"
 arc-swap = "1.1"
 async-trait = "0.1.7"
-err-derive = "0.2.3"
-bytes = "0.6.0"
-lru = { version = "0.6", optional = true }
+err-derive = "0.3"
+bytes = "1.2"
+lru = { version = "0.7", optional = true }
+cfg-if = "1.0"
 
 sodiumoxide = { version = "0.2.5-0", package = "kuska-sodiumoxide" }
-kuska-handshake = { version = "0.2.0", git = "https://github.com/Alexis211/handshake", branch = "tokio1.0", features = ["default", "async_std"] }
+kuska-handshake = { version = "0.2.0", features = ["default", "async_std"] }
+
+opentelemetry = { version = "0.17", optional = true }
+opentelemetry-contrib = { version = "0.9", optional = true }
 
 [dev-dependencies]
+env_logger = "0.9"
 structopt = { version = "0.3", default-features = false }
 chrono = "0.4"
 

Makefile

@@ -1,5 +1,10 @@
 all:
+	#cargo build --all-features
 	cargo build
 	cargo build --example fullmesh
-	#RUST_LOG=netapp=debug cargo run --example fullmesh -- -n 3242ce79e05e8b6a0e43441fbd140a906e13f335f298ae3a52f29784abbab500 -p 6c304114a0e1018bbe60502a34d33f4f439f370856c3333dda2726da01eb93a4894b7ef7249a71f11d342b69702f1beb7c93ec95fbcf122ad1eca583bb0629e7
+	cargo build --all-features --example basalt
+	RUST_LOG=netapp=trace,fullmesh=trace cargo run --example fullmesh -- -n 3242ce79e05e8b6a0e43441fbd140a906e13f335f298ae3a52f29784abbab500 -p 6c304114a0e1018bbe60502a34d33f4f439f370856c3333dda2726da01eb93a4894b7ef7249a71f11d342b69702f1beb7c93ec95fbcf122ad1eca583bb0629e7
+	#RUST_LOG=netapp=debug,fullmesh=debug cargo run --example fullmesh
 
+test:
+	cargo test --all-features -- --test-threads 1

examples/basalt.rs

@@ -1,20 +1,23 @@
 use std::io::Write;
-use std::net::SocketAddr;
 use std::sync::Arc;
 use std::time::Duration;
 
 use log::{debug, info, warn};
 
+use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 use structopt::StructOpt;
 
 use sodiumoxide::crypto::auth;
 use sodiumoxide::crypto::sign::ed25519;
 
+use tokio::sync::watch;
+
+use netapp::endpoint::*;
 use netapp::message::*;
 use netapp::peering::basalt::*;
-use netapp::proto::*;
-use netapp::NetApp;
+use netapp::util::parse_peer_addr;
+use netapp::{NetApp, NodeID};
 
 #[derive(StructOpt, Debug)]
 #[structopt(name = "netapp")]
@@ -50,6 +53,12 @@ pub struct Opt {
 	reset_count: usize,
 }
 
+struct Example {
+	netapp: Arc<NetApp>,
+	basalt: Arc<Basalt>,
+	example_endpoint: Arc<Endpoint<ExampleMessage, Self>>,
+}
+
 #[tokio::main]
 async fn main() {
 	env_logger::Builder::new()
@@ -85,16 +94,11 @@ async fn main() {
 	info!("KYEV SK {}", hex::encode(&privkey));
 	info!("KYEV PK {}", hex::encode(&privkey.public_key()));
 
-	let netapp = NetApp::new(netid, privkey);
+	let netapp = NetApp::new(0u64, netid, privkey);
 
 	let mut bootstrap_peers = vec![];
 	for peer in opt.bootstrap_peers.iter() {
-		if let Some(delim) = peer.find('@') {
-			let (key, ip) = peer.split_at(delim);
-			let pubkey = ed25519::PublicKey::from_slice(&hex::decode(&key).unwrap()).unwrap();
-			let ip = ip[1..].parse::<SocketAddr>().unwrap();
-			bootstrap_peers.push((pubkey, ip));
-		}
+		bootstrap_peers.push(parse_peer_addr(peer).expect("Invalid peer address"));
 	}
 
 	let basalt_params = BasaltParams {
@@ -104,40 +108,44 @@ async fn main() {
 		reset_interval: Duration::from_secs(opt.reset_interval),
 		reset_count: opt.reset_count,
 	};
-	let peering = Basalt::new(netapp.clone(), bootstrap_peers, basalt_params);
+	let basalt = Basalt::new(netapp.clone(), bootstrap_peers, basalt_params);
 
-	netapp.add_msg_handler::<ExampleMessage, _, _>(
-		|_from: ed25519::PublicKey, msg: ExampleMessage| {
-			debug!("Got example message: {:?}, sending example response", msg);
-			async {
-				ExampleResponse {
-					example_field: false,
-				}
-			}
-		},
-	);
+	let example = Arc::new(Example {
+		netapp: netapp.clone(),
+		basalt,
+		example_endpoint: netapp.endpoint("__netapp/examples/basalt.rs/Example".into()),
+	});
+	example.example_endpoint.set_handler(example.clone());
 
 	let listen_addr = opt.listen_addr.parse().unwrap();
 	let public_addr = opt.public_addr.map(|x| x.parse().unwrap());
+
+	let watch_cancel = netapp::util::watch_ctrl_c();
+
 	tokio::join!(
-		sampling_loop(netapp.clone(), peering.clone()),
-		netapp.listen(listen_addr, public_addr),
-		peering.run(),
+		example.clone().sampling_loop(watch_cancel.clone()),
+		example
+			.netapp
+			.clone()
+			.listen(listen_addr, public_addr, watch_cancel.clone()),
+		example.basalt.clone().run(watch_cancel.clone()),
 	);
 }
 
-async fn sampling_loop(netapp: Arc<NetApp>, basalt: Arc<Basalt>) {
-	loop {
+impl Example {
+	async fn sampling_loop(self: Arc<Self>, must_exit: watch::Receiver<bool>) {
+		while !*must_exit.borrow() {
 			tokio::time::sleep(Duration::from_secs(10)).await;
 
-		let peers = basalt.sample(10);
+			let peers = self.basalt.sample(10);
 			for p in peers {
 				debug!("kyev S {}", hex::encode(p));
 
-			let netapp2 = netapp.clone();
+				let self2 = self.clone();
 				tokio::spawn(async move {
-				match netapp2
-					.request(&p, ExampleMessage { example_field: 42 }, PRIO_NORMAL)
+					match self2
+						.example_endpoint
+						.call(&p, ExampleMessage { example_field: 42 }, PRIO_NORMAL)
 						.await
 					{
 						Ok(resp) => debug!("Got example response: {:?}", resp),
@@ -146,6 +154,17 @@ async fn sampling_loop(netapp: Arc<NetApp>, basalt: Arc<Basalt>) {
 				});
 			}
 		}
+	}
+}
+
+#[async_trait]
+impl EndpointHandler<ExampleMessage> for Example {
+	async fn handle(self: &Arc<Self>, msg: &ExampleMessage, _from: NodeID) -> ExampleResponse {
+		debug!("Got example message: {:?}, sending example response", msg);
+		ExampleResponse {
+			example_field: false,
+		}
+	}
 }
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -159,6 +178,5 @@ struct ExampleResponse {
 }
 
 impl Message for ExampleMessage {
-	const KIND: MessageKind = 0x99000001;
 	type Response = ExampleResponse;
 }

examples/fullmesh.rs

@@ -1,16 +1,24 @@
 use std::io::Write;
 use std::net::SocketAddr;
+use std::sync::Arc;
+use std::time::Duration;
 
-use log::info;
-
+use async_trait::async_trait;
+use bytes::Bytes;
+use futures::{stream, StreamExt};
+use log::*;
+use serde::{Deserialize, Serialize};
 use structopt::StructOpt;
+use tokio::sync::watch;
 
 use sodiumoxide::crypto::auth;
 use sodiumoxide::crypto::sign::ed25519;
 
+use netapp::endpoint::*;
+use netapp::message::*;
 use netapp::peering::fullmesh::*;
-use netapp::NetApp;
-use netapp::NodeID;
+use netapp::util::*;
+use netapp::{NetApp, NodeID};
 
 #[derive(StructOpt, Debug)]
 #[structopt(name = "netapp")]
@@ -64,23 +72,148 @@ async fn main() {
 	};
 
 	info!("Node private key: {}", hex::encode(&privkey));
-	info!("Node public key: {}", hex::encode(&privkey.public_key()));
+	info!("Node public key: {}", hex::encode(privkey.public_key()));
 
-	let netapp = NetApp::new(netid, privkey);
+	let public_addr = opt.public_addr.map(|x| x.parse().unwrap());
+	let listen_addr: SocketAddr = opt.listen_addr.parse().unwrap();
+	info!("Node public address: {:?}", public_addr);
+	info!("Node listen address: {}", listen_addr);
+
+	let netapp = NetApp::new(0u64, netid.clone(), privkey.clone());
 
 	let mut bootstrap_peers = vec![];
 	for peer in opt.bootstrap_peers.iter() {
-		if let Some(delim) = peer.find('@') {
-			let (key, ip) = peer.split_at(delim);
-			let pubkey = NodeID::from_slice(&hex::decode(&key).unwrap()).unwrap();
-			let ip = ip[1..].parse::<SocketAddr>().unwrap();
-			bootstrap_peers.push((pubkey, ip));
-		}
+		bootstrap_peers.push(parse_peer_addr(peer).expect("Invalid peer address"));
 	}
 
-	let peering = FullMeshPeeringStrategy::new(netapp.clone(), bootstrap_peers);
+	let peering = FullMeshPeeringStrategy::new(
+		netapp.clone(),
+		bootstrap_peers,
+		public_addr.map(|a| SocketAddr::new(a, listen_addr.port())),
+	);
 
-	let listen_addr = opt.listen_addr.parse().unwrap();
-	let public_addr = opt.public_addr.map(|x| x.parse().unwrap());
-	tokio::join!(netapp.listen(listen_addr, public_addr), peering.run(),);
+	info!("Add more peers to this mesh by running: fullmesh -n {} -l 127.0.0.1:$((1000 + $RANDOM)) -b {}@{}",
+		hex::encode(&netid),
+		hex::encode(privkey.public_key()),
+		listen_addr);
+
+	let watch_cancel = netapp::util::watch_ctrl_c();
+
+	let example = Arc::new(Example {
+		netapp: netapp.clone(),
+		fullmesh: peering.clone(),
+		example_endpoint: netapp.endpoint("__netapp/examples/fullmesh.rs/Example".into()),
+	});
+	example.example_endpoint.set_handler(example.clone());
+
+	tokio::join!(
+		example.exchange_loop(watch_cancel.clone()),
+		netapp.listen(listen_addr, public_addr, watch_cancel.clone()),
+		peering.run(watch_cancel),
+	);
+}
+
+// ----
+
+struct Example {
+	netapp: Arc<NetApp>,
+	fullmesh: Arc<FullMeshPeeringStrategy>,
+	example_endpoint: Arc<Endpoint<ExampleMessage, Self>>,
+}
+
+impl Example {
+	async fn exchange_loop(self: Arc<Self>, must_exit: watch::Receiver<bool>) {
+		let mut i = 12000;
+		while !*must_exit.borrow() {
+			tokio::time::sleep(Duration::from_secs(2)).await;
+
+			let peers = self.fullmesh.get_peer_list();
+			for p in peers.iter() {
+				let id = p.id;
+				if id == self.netapp.id {
+					continue;
+				}
+				i += 1;
+				let example_field = i;
+				let self2 = self.clone();
+				tokio::spawn(async move {
+					info!(
+						"Send example query {} to {}",
+						example_field,
+						hex::encode(id)
+					);
+					// Fake data stream with some delays in item production
+					let stream =
+						Box::pin(stream::iter([100, 200, 300, 400]).then(|x| async move {
+							tokio::time::sleep(Duration::from_millis(500)).await;
+							Ok(Bytes::from(vec![(x % 256) as u8; 133 * x]))
+						}));
+					match self2
+						.example_endpoint
+						.call_streaming(
+							&id,
+							Req::new(ExampleMessage { example_field })
+								.unwrap()
+								.with_stream(stream),
+							PRIO_NORMAL,
+						)
+						.await
+					{
+						Ok(resp) => {
+							let (resp, stream) = resp.into_parts();
+							info!(
+								"Got example response to {} from {}: {:?}",
+								example_field,
+								hex::encode(id),
+								resp
+							);
+							let mut stream = stream.unwrap();
+							while let Some(x) = stream.next().await {
+								info!("Response: stream got bytes {:?}", x.map(|b| b.len()));
+							}
+						}
+						Err(e) => warn!("Error with example request: {}", e),
+					}
+				});
+			}
+		}
+	}
+}
+
+#[async_trait]
+impl StreamingEndpointHandler<ExampleMessage> for Example {
+	async fn handle(
+		self: &Arc<Self>,
+		mut msg: Req<ExampleMessage>,
+		_from: NodeID,
+	) -> Resp<ExampleMessage> {
+		info!(
+			"Got example message: {:?}, sending example response",
+			msg.msg()
+		);
+		let source_stream = msg.take_stream().unwrap();
+		// Return same stream with 300ms delay
+		let new_stream = Box::pin(source_stream.then(|x| async move {
+			tokio::time::sleep(Duration::from_millis(300)).await;
+			x
+		}));
+		Resp::new(ExampleResponse {
+			example_field: false,
+		})
+		.with_stream(new_stream)
+	}
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+struct ExampleMessage {
+	example_field: usize,
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+struct ExampleResponse {
+	example_field: bool,
+}
+
+impl Message for ExampleMessage {
+	type Response = ExampleResponse;
 }

src/bytes_buf.rs

@@ -0,0 +1,186 @@
+use std::cmp::Ordering;
+use std::collections::VecDeque;
+
+use bytes::BytesMut;
+
+pub use bytes::Bytes;
+
+/// A circular buffer of bytes, internally represented as a list of Bytes
+/// for optimization, but that for all intent and purposes acts just like
+/// a big byte slice which can be extended on the right and from which
+/// stuff can be taken on the left.
+pub struct BytesBuf {
+	buf: VecDeque<Bytes>,
+	buf_len: usize,
+}
+
+impl BytesBuf {
+	/// Creates a new empty BytesBuf
+	pub fn new() -> Self {
+		Self {
+			buf: VecDeque::new(),
+			buf_len: 0,
+		}
+	}
+
+	/// Returns the number of bytes stored in the BytesBuf
+	#[inline]
+	pub fn len(&self) -> usize {
+		self.buf_len
+	}
+
+	/// Returns true iff the BytesBuf contains zero bytes
+	#[inline]
+	pub fn is_empty(&self) -> bool {
+		self.buf_len == 0
+	}
+
+	/// Adds some bytes to the right of the buffer
+	pub fn extend(&mut self, b: Bytes) {
+		if !b.is_empty() {
+			self.buf_len += b.len();
+			self.buf.push_back(b);
+		}
+	}
+
+	/// Takes the whole content of the buffer and returns it as a single Bytes unit
+	pub fn take_all(&mut self) -> Bytes {
+		if self.buf.is_empty() {
+			Bytes::new()
+		} else if self.buf.len() == 1 {
+			self.buf_len = 0;
+			self.buf.pop_back().unwrap()
+		} else {
+			let mut ret = BytesMut::with_capacity(self.buf_len);
+			for b in self.buf.iter() {
+				ret.extend_from_slice(&b[..]);
+			}
+			self.buf.clear();
+			self.buf_len = 0;
+			ret.freeze()
+		}
+	}
+
+	/// Takes at most max_len bytes from the left of the buffer
+	pub fn take_max(&mut self, max_len: usize) -> Bytes {
+		if self.buf_len <= max_len {
+			self.take_all()
+		} else {
+			self.take_exact_ok(max_len)
+		}
+	}
+
+	/// Take exactly len bytes from the left of the buffer, returns None if
+	/// the BytesBuf doesn't contain enough data
+	pub fn take_exact(&mut self, len: usize) -> Option<Bytes> {
+		if self.buf_len < len {
+			None
+		} else {
+			Some(self.take_exact_ok(len))
+		}
+	}
+
+	fn take_exact_ok(&mut self, len: usize) -> Bytes {
+		assert!(len <= self.buf_len);
+		let front = self.buf.pop_front().unwrap();
+		match front.len().cmp(&len) {
+			Ordering::Greater => {
+				self.buf.push_front(front.slice(len..));
+				self.buf_len -= len;
+				front.slice(..len)
+			}
+			Ordering::Equal => {
+				self.buf_len -= len;
+				front
+			}
+			Ordering::Less => {
+				let mut ret = BytesMut::with_capacity(len);
+				ret.extend_from_slice(&front[..]);
+				self.buf_len -= front.len();
+				while ret.len() < len {
+					let front = self.buf.pop_front().unwrap();
+					if front.len() > len - ret.len() {
+						let take = len - ret.len();
+						ret.extend_from_slice(&front[..take]);
+						self.buf.push_front(front.slice(take..));
+						self.buf_len -= take;
+						break;
+					} else {
+						ret.extend_from_slice(&front[..]);
+						self.buf_len -= front.len();
+					}
+				}
+				ret.freeze()
+			}
+		}
+	}
+
+	/// Return the internal sequence of Bytes slices that make up the buffer
+	pub fn into_slices(self) -> VecDeque<Bytes> {
+		self.buf
+	}
+}
+
+impl Default for BytesBuf {
+	fn default() -> Self {
+		Self::new()
+	}
+}
+
+impl From<Bytes> for BytesBuf {
+	fn from(b: Bytes) -> BytesBuf {
+		let mut ret = BytesBuf::new();
+		ret.extend(b);
+		ret
+	}
+}
+
+impl From<BytesBuf> for Bytes {
+	fn from(mut b: BytesBuf) -> Bytes {
+		b.take_all()
+	}
+}
+
+#[cfg(test)]
+mod test {
+	use super::*;
+
+	#[test]
+	fn test_bytes_buf() {
+		let mut buf = BytesBuf::new();
+		assert!(buf.len() == 0);
+		assert!(buf.is_empty());
+
+		buf.extend(Bytes::from(b"Hello, world!".to_vec()));
+		assert!(buf.len() == 13);
+		assert!(!buf.is_empty());
+
+		buf.extend(Bytes::from(b"1234567890".to_vec()));
+		assert!(buf.len() == 23);
+		assert!(!buf.is_empty());
+
+		assert_eq!(
+			buf.take_all(),
+			Bytes::from(b"Hello, world!1234567890".to_vec())
+		);
+		assert!(buf.len() == 0);
+		assert!(buf.is_empty());
+
+		buf.extend(Bytes::from(b"1234567890".to_vec()));
+		buf.extend(Bytes::from(b"Hello, world!".to_vec()));
+		assert!(buf.len() == 23);
+		assert!(!buf.is_empty());
+
+		assert_eq!(buf.take_max(12), Bytes::from(b"1234567890He".to_vec()));
+		assert!(buf.len() == 11);
+
+		assert_eq!(buf.take_exact(12), None);
+		assert!(buf.len() == 11);
+		assert_eq!(
+			buf.take_exact(11),
+			Some(Bytes::from(b"llo, world!".to_vec()))
+		);
+		assert!(buf.len() == 0);
+		assert!(buf.is_empty());
+	}
+}

src/client.rs

@@ -0,0 +1,292 @@
+use std::collections::HashMap;
+use std::net::SocketAddr;
+use std::pin::Pin;
+use std::sync::atomic::{self, AtomicU32};
+use std::sync::{Arc, Mutex};
+use std::task::Poll;
+
+use arc_swap::ArcSwapOption;
+use async_trait::async_trait;
+use bytes::Bytes;
+use log::{debug, error, trace};
+
+use futures::io::AsyncReadExt;
+use futures::Stream;
+use kuska_handshake::async_std::{handshake_client, BoxStream};
+use tokio::net::TcpStream;
+use tokio::select;
+use tokio::sync::{mpsc, oneshot, watch};
+use tokio_util::compat::*;
+
+#[cfg(feature = "telemetry")]
+use opentelemetry::{
+	trace::{FutureExt, Span, SpanKind, TraceContextExt, Tracer},
+	Context, KeyValue,
+};
+#[cfg(feature = "telemetry")]
+use opentelemetry_contrib::trace::propagator::binary::*;
+
+use crate::error::*;
+use crate::message::*;
+use crate::netapp::*;
+use crate::recv::*;
+use crate::send::*;
+use crate::stream::*;
+use crate::util::*;
+
+pub(crate) struct ClientConn {
+	pub(crate) remote_addr: SocketAddr,
+	pub(crate) peer_id: NodeID,
+
+	query_send: ArcSwapOption<mpsc::UnboundedSender<SendItem>>,
+
+	next_query_number: AtomicU32,
+	inflight: Mutex<HashMap<RequestID, oneshot::Sender<ByteStream>>>,
+}
+
+impl ClientConn {
+	pub(crate) async fn init(
+		netapp: Arc<NetApp>,
+		socket: TcpStream,
+		peer_id: NodeID,
+	) -> Result<(), Error> {
+		let remote_addr = socket.peer_addr()?;
+		let mut socket = socket.compat();
+
+		// Do handshake to authenticate and prove our identity to server
+		let handshake = handshake_client(
+			&mut socket,
+			netapp.netid.clone(),
+			netapp.id,
+			netapp.privkey.clone(),
+			peer_id,
+		)
+		.await?;
+
+		debug!(
+			"Handshake complete (client) with {}@{}",
+			hex::encode(peer_id),
+			remote_addr
+		);
+
+		// Create BoxStream layer that encodes content
+		let (read, write) = socket.split();
+		let (mut read, write) =
+			BoxStream::from_handshake(read, write, handshake, 0x8000).split_read_write();
+
+		// Before doing anything, receive version tag and
+		// check they are running the same version as us
+		let mut their_version_tag = VersionTag::default();
+		read.read_exact(&mut their_version_tag[..]).await?;
+		if their_version_tag != netapp.version_tag {
+			let msg = format!(
+				"different version tags: {} (theirs) vs. {} (ours)",
+				hex::encode(their_version_tag),
+				hex::encode(netapp.version_tag)
+			);
+			error!("Cannot connect to {}: {}", hex::encode(&peer_id[..8]), msg);
+			return Err(Error::VersionMismatch(msg));
+		}
+
+		// Build and launch stuff that manages sending requests client-side
+		let (query_send, query_recv) = mpsc::unbounded_channel();
+
+		let (stop_recv_loop, stop_recv_loop_recv) = watch::channel(false);
+
+		let conn = Arc::new(ClientConn {
+			remote_addr,
+			peer_id,
+			next_query_number: AtomicU32::from(RequestID::default()),
+			query_send: ArcSwapOption::new(Some(Arc::new(query_send))),
+			inflight: Mutex::new(HashMap::new()),
+		});
+
+		netapp.connected_as_client(peer_id, conn.clone());
+
+		let debug_name = format!("CLI {}", hex::encode(&peer_id[..8]));
+
+		tokio::spawn(async move {
+			let debug_name_2 = debug_name.clone();
+			let send_future = tokio::spawn(conn.clone().send_loop(query_recv, write, debug_name_2));
+
+			let conn2 = conn.clone();
+			let recv_future = tokio::spawn(async move {
+				select! {
+					r = conn2.recv_loop(read, debug_name) => r,
+					_ = await_exit(stop_recv_loop_recv) => Ok(())
+				}
+			});
+
+			send_future.await.log_err("ClientConn send_loop");
+
+			// FIXME: should do here: wait for inflight requests to all have their response
+			stop_recv_loop
+				.send(true)
+				.log_err("ClientConn send true to stop_recv_loop");
+
+			recv_future.await.log_err("ClientConn recv_loop");
+
+			// Make sure we don't wait on any more requests that won't
+			// have a response
+			conn.inflight.lock().unwrap().clear();
+
+			netapp.disconnected_as_client(&peer_id, conn);
+		});
+
+		Ok(())
+	}
+
+	pub fn close(&self) {
+		self.query_send.store(None);
+	}
+
+	pub(crate) async fn call<T>(
+		self: Arc<Self>,
+		req: Req<T>,
+		path: &str,
+		prio: RequestPriority,
+	) -> Result<Resp<T>, Error>
+	where
+		T: Message,
+	{
+		let query_send = self.query_send.load_full().ok_or(Error::ConnectionClosed)?;
+
+		let id = self
+			.next_query_number
+			.fetch_add(1, atomic::Ordering::Relaxed);
+
+		cfg_if::cfg_if! {
+			if #[cfg(feature = "telemetry")] {
+				let tracer = opentelemetry::global::tracer("netapp");
+				let mut span = tracer.span_builder(format!("RPC >> {}", path))
+					.with_kind(SpanKind::Client)
+					.start(&tracer);
+				let propagator = BinaryPropagator::new();
+				let telemetry_id: Bytes = propagator.to_bytes(span.span_context()).to_vec().into();
+			} else {
+				let telemetry_id: Bytes = Bytes::new();
+			}
+		};
+
+		// Encode request
+		let req_enc = req.into_enc(prio, path.as_bytes().to_vec().into(), telemetry_id);
+		let req_msg_len = req_enc.msg.len();
+		let (req_stream, req_order) = req_enc.encode();
+
+		// Send request through
+		let (resp_send, resp_recv) = oneshot::channel();
+		let old = self.inflight.lock().unwrap().insert(id, resp_send);
+		if let Some(old_ch) = old {
+			error!(
+				"Too many inflight requests! RequestID collision. Interrupting previous request."
+			);
+			let _ = old_ch.send(Box::pin(futures::stream::once(async move {
+				Err(std::io::Error::new(
+					std::io::ErrorKind::Other,
+					"RequestID collision, too many inflight requests",
+				))
+			})));
+		}
+
+		debug!(
+			"request: query_send {}, path {}, prio {} (serialized message: {} bytes)",
+			id, path, prio, req_msg_len
+		);
+
+		#[cfg(feature = "telemetry")]
+		span.set_attribute(KeyValue::new("len_query_msg", req_msg_len as i64));
+
+		query_send.send(SendItem::Stream(id, prio, req_order, req_stream))?;
+
+		let canceller = CancelOnDrop::new(id, query_send.as_ref().clone());
+
+		cfg_if::cfg_if! {
+			if #[cfg(feature = "telemetry")] {
+				let stream = resp_recv
+					.with_context(Context::current_with_span(span))
+					.await?;
+			} else {
+				let stream = resp_recv.await?;
+			}
+		}
+
+		let stream = Box::pin(canceller.for_stream(stream));
+
+		let resp_enc = RespEnc::decode(stream).await?;
+		debug!("client: got response to request {} (path {})", id, path);
+		Resp::from_enc(resp_enc)
+	}
+}
+
+impl SendLoop for ClientConn {}
+
+#[async_trait]
+impl RecvLoop for ClientConn {
+	fn recv_handler(self: &Arc<Self>, id: RequestID, stream: ByteStream) {
+		trace!("ClientConn recv_handler {}", id);
+
+		let mut inflight = self.inflight.lock().unwrap();
+		if let Some(ch) = inflight.remove(&id) {
+			if ch.send(stream).is_err() {
+				debug!("Could not send request response, probably because request was interrupted. Dropping response.");
+			}
+		} else {
+			debug!("Got unexpected response to request {}, dropping it", id);
+		}
+	}
+}
+
+// ----
+
+struct CancelOnDrop {
+	id: RequestID,
+	query_send: mpsc::UnboundedSender<SendItem>,
+}
+
+impl CancelOnDrop {
+	fn new(id: RequestID, query_send: mpsc::UnboundedSender<SendItem>) -> Self {
+		Self { id, query_send }
+	}
+	fn for_stream(self, stream: ByteStream) -> CancelOnDropStream {
+		CancelOnDropStream {
+			cancel: Some(self),
+			stream,
+		}
+	}
+}
+
+impl Drop for CancelOnDrop {
+	fn drop(&mut self) {
+		trace!("cancelling request {}", self.id);
+		let _ = self.query_send.send(SendItem::Cancel(self.id));
+	}
+}
+
+#[pin_project::pin_project]
+struct CancelOnDropStream {
+	cancel: Option<CancelOnDrop>,
+	#[pin]
+	stream: ByteStream,
+}
+
+impl Stream for CancelOnDropStream {
+	type Item = Packet;
+
+	fn poll_next(
+		self: Pin<&mut Self>,
+		cx: &mut std::task::Context<'_>,
+	) -> Poll<Option<Self::Item>> {
+		let this = self.project();
+		let res = this.stream.poll_next(cx);
+		if matches!(res, Poll::Ready(None)) {
+			if let Some(c) = this.cancel.take() {
+				std::mem::forget(c)
+			}
+		}
+		res
+	}
+
+	fn size_hint(&self) -> (usize, Option<usize>) {
+		self.stream.size_hint()
+	}
+}

src/conn.rs

@@ -1,271 +0,0 @@
-use std::collections::HashMap;
-use std::net::SocketAddr;
-use std::sync::atomic::{self, AtomicBool, AtomicU16};
-use std::sync::{Arc, Mutex};
-
-use bytes::Bytes;
-use log::{debug, error, trace};
-
-use tokio::net::TcpStream;
-use tokio::sync::{mpsc, oneshot, watch};
-use tokio_util::compat::*;
-
-use futures::io::AsyncReadExt;
-
-use async_trait::async_trait;
-
-use kuska_handshake::async_std::{handshake_client, handshake_server, BoxStream};
-
-use crate::error::*;
-use crate::message::*;
-use crate::netapp::*;
-use crate::proto::*;
-use crate::util::*;
-
-pub(crate) struct ServerConn {
-	pub(crate) remote_addr: SocketAddr,
-	pub(crate) peer_id: NodeID,
-
-	netapp: Arc<NetApp>,
-
-	resp_send: mpsc::UnboundedSender<Option<(RequestID, RequestPriority, Vec<u8>)>>,
-	close_send: watch::Sender<bool>,
-}
-
-impl ServerConn {
-	pub(crate) async fn run(netapp: Arc<NetApp>, socket: TcpStream) -> Result<(), Error> {
-		let remote_addr = socket.peer_addr()?;
-		let mut socket = socket.compat();
-
-		let handshake = handshake_server(
-			&mut socket,
-			netapp.netid.clone(),
-			netapp.id,
-			netapp.privkey.clone(),
-		)
-		.await?;
-		let peer_id = handshake.peer_pk;
-
-		debug!(
-			"Handshake complete (server) with {}@{}",
-			hex::encode(&peer_id),
-			remote_addr
-		);
-
-		let (read, write) = socket.split();
-
-		let (read, write) =
-			BoxStream::from_handshake(read, write, handshake, 0x8000).split_read_write();
-
-		let (resp_send, resp_recv) = mpsc::unbounded_channel();
-
-		let (close_send, close_recv) = watch::channel(false);
-
-		let conn = Arc::new(ServerConn {
-			netapp: netapp.clone(),
-			remote_addr,
-			peer_id,
-			resp_send,
-			close_send,
-		});
-
-		netapp.connected_as_server(peer_id, conn.clone());
-
-		let conn2 = conn.clone();
-		let conn3 = conn.clone();
-		let close_recv2 = close_recv.clone();
-		tokio::try_join!(
-			async move {
-				tokio::select!(
-					r = conn2.recv_loop(read) => r,
-					_ = await_exit(close_recv) => Ok(()),
-				)
-			},
-			async move {
-				tokio::select!(
-					r = conn3.send_loop(resp_recv, write) => r,
-					_ = await_exit(close_recv2) => Ok(()),
-				)
-			},
-		)
-		.map(|_| ())
-		.log_err("ServerConn recv_loop/send_loop");
-
-		netapp.disconnected_as_server(&peer_id, conn);
-
-		Ok(())
-	}
-
-	pub fn close(&self) {
-		self.close_send.send(true).unwrap();
-	}
-}
-
-impl SendLoop for ServerConn {}
-
-#[async_trait]
-impl RecvLoop for ServerConn {
-	async fn recv_handler(self: Arc<Self>, id: u16, bytes: Vec<u8>) {
-		trace!("ServerConn recv_handler {} ({} bytes)", id, bytes.len());
-
-		let bytes: Bytes = bytes.into();
-
-		let prio = bytes[0];
-
-		let mut kind_bytes = [0u8; 4];
-		kind_bytes.copy_from_slice(&bytes[1..5]);
-		let kind = u32::from_be_bytes(kind_bytes);
-
-		if let Some(handler) = self.netapp.msg_handlers.load().get(&kind) {
-			let net_handler = &handler.net_handler;
-			let resp = net_handler(self.peer_id, bytes.slice(5..)).await;
-			self.resp_send
-				.send(Some((id, prio, resp)))
-				.log_err("ServerConn recv_handler send resp");
-		}
-	}
-}
-pub(crate) struct ClientConn {
-	pub(crate) remote_addr: SocketAddr,
-	pub(crate) peer_id: NodeID,
-
-	query_send: mpsc::UnboundedSender<Option<(RequestID, RequestPriority, Vec<u8>)>>,
-
-	next_query_number: AtomicU16,
-	inflight: Mutex<HashMap<RequestID, oneshot::Sender<Vec<u8>>>>,
-	must_exit: AtomicBool,
-	stop_recv_loop: watch::Sender<bool>,
-}
-
-impl ClientConn {
-	pub(crate) async fn init(
-		netapp: Arc<NetApp>,
-		socket: TcpStream,
-		peer_id: NodeID,
-	) -> Result<(), Error> {
-		let remote_addr = socket.peer_addr()?;
-		let mut socket = socket.compat();
-
-		let handshake = handshake_client(
-			&mut socket,
-			netapp.netid.clone(),
-			netapp.id,
-			netapp.privkey.clone(),
-			peer_id,
-		)
-		.await?;
-
-		debug!(
-			"Handshake complete (client) with {}@{}",
-			hex::encode(&peer_id),
-			remote_addr
-		);
-
-		let (read, write) = socket.split();
-
-		let (read, write) =
-			BoxStream::from_handshake(read, write, handshake, 0x8000).split_read_write();
-
-		let (query_send, query_recv) = mpsc::unbounded_channel();
-
-		let (stop_recv_loop, stop_recv_loop_recv) = watch::channel(false);
-
-		let conn = Arc::new(ClientConn {
-			remote_addr,
-			peer_id,
-			next_query_number: AtomicU16::from(0u16),
-			query_send,
-			inflight: Mutex::new(HashMap::new()),
-			must_exit: AtomicBool::new(false),
-			stop_recv_loop,
-		});
-
-		netapp.connected_as_client(peer_id, conn.clone());
-
-		tokio::spawn(async move {
-			let conn2 = conn.clone();
-			let conn3 = conn.clone();
-			tokio::try_join!(conn2.send_loop(query_recv, write), async move {
-				tokio::select!(
-					r = conn3.recv_loop(read) => r,
-					_ = await_exit(stop_recv_loop_recv) => Ok(()),
-				)
-			})
-			.map(|_| ())
-			.log_err("ClientConn send_loop/recv_loop/dispatch_loop");
-
-			netapp.disconnected_as_client(&peer_id, conn);
-		});
-
-		Ok(())
-	}
-
-	pub fn close(&self) {
-		self.must_exit.store(true, atomic::Ordering::SeqCst);
-		self.query_send
-			.send(None)
-			.log_err("could not write None in query_send");
-		if self.inflight.lock().unwrap().is_empty() {
-			self.stop_recv_loop
-				.send(true)
-				.log_err("could not write true to stop_recv_loop");
-		}
-	}
-
-	pub(crate) async fn request<T>(
-		self: Arc<Self>,
-		rq: T,
-		prio: RequestPriority,
-	) -> Result<<T as Message>::Response, Error>
-	where
-		T: Message,
-	{
-		let id = self
-			.next_query_number
-			.fetch_add(1u16, atomic::Ordering::Relaxed);
-		let mut bytes = vec![prio];
-		bytes.extend_from_slice(&u32::to_be_bytes(T::KIND)[..]);
-		bytes.extend_from_slice(&rmp_to_vec_all_named(&rq)?[..]);
-
-		let (resp_send, resp_recv) = oneshot::channel();
-		let old = self.inflight.lock().unwrap().insert(id, resp_send);
-		if let Some(old_ch) = old {
-			error!(
-				"Too many inflight requests! RequestID collision. Interrupting previous request."
-			);
-			if old_ch.send(vec![]).is_err() {
-				debug!("Could not send empty response to collisionned request, probably because request was interrupted. Dropping response.");
-			}
-		}
-
-		trace!("request: query_send {}, {} bytes", id, bytes.len());
-		self.query_send.send(Some((id, prio, bytes)))?;
-
-		let resp = resp_recv.await?;
-
-		rmp_serde::decode::from_read_ref::<_, Result<<T as Message>::Response, String>>(&resp[..])?
-			.map_err(Error::Remote)
-	}
-}
-
-impl SendLoop for ClientConn {}
-
-#[async_trait]
-impl RecvLoop for ClientConn {
-	async fn recv_handler(self: Arc<Self>, id: RequestID, msg: Vec<u8>) {
-		trace!("ClientConn recv_handler {} ({} bytes)", id, msg.len());
-
-		let mut inflight = self.inflight.lock().unwrap();
-		if let Some(ch) = inflight.remove(&id) {
-			if ch.send(msg).is_err() {
-				debug!("Could not send request response, probably because request was interrupted. Dropping response.");
-			}
-		}
-
-		if inflight.is_empty() && self.must_exit.load(atomic::Ordering::SeqCst) {
-			self.stop_recv_loop
-				.send(true)
-				.log_err("could not write true to stop_recv_loop");
-		}
-	}
-}

src/endpoint.rs

@@ -0,0 +1,201 @@
+use std::marker::PhantomData;
+use std::sync::Arc;
+
+use arc_swap::ArcSwapOption;
+use async_trait::async_trait;
+
+use crate::error::Error;
+use crate::message::*;
+use crate::netapp::*;
+
+/// This trait should be implemented by an object of your application
+/// that can handle a message of type `M`, if it wishes to handle
+/// streams attached to the request and/or to send back streams
+/// attached to the response..
+///
+/// The handler object should be in an Arc, see `Endpoint::set_handler`
+#[async_trait]
+pub trait StreamingEndpointHandler<M>: Send + Sync
+where
+	M: Message,
+{
+	async fn handle(self: &Arc<Self>, m: Req<M>, from: NodeID) -> Resp<M>;
+}
+
+/// If one simply wants to use an endpoint in a client fashion,
+/// without locally serving requests to that endpoint,
+/// use the unit type `()` as the handler type:
+/// it will panic if it is ever made to handle request.
+#[async_trait]
+impl<M: Message> EndpointHandler<M> for () {
+	async fn handle(self: &Arc<()>, _m: &M, _from: NodeID) -> M::Response {
+		panic!("This endpoint should not have a local handler.");
+	}
+}
+
+// ----
+
+/// This trait should be implemented by an object of your application
+/// that can handle a message of type `M`, in the cases where it doesn't
+/// care about attached stream in the request nor in the response.
+#[async_trait]
+pub trait EndpointHandler<M>: Send + Sync
+where
+	M: Message,
+{
+	async fn handle(self: &Arc<Self>, m: &M, from: NodeID) -> M::Response;
+}
+
+#[async_trait]
+impl<T, M> StreamingEndpointHandler<M> for T
+where
+	T: EndpointHandler<M>,
+	M: Message,
+{
+	async fn handle(self: &Arc<Self>, mut m: Req<M>, from: NodeID) -> Resp<M> {
+		// Immediately drop stream to ignore all data that comes in,
+		// instead of buffering it indefinitely
+		drop(m.take_stream());
+		Resp::new(EndpointHandler::handle(self, m.msg(), from).await)
+	}
+}
+
+// ----
+
+/// This struct represents an endpoint for message of type `M`.
+///
+/// Creating a new endpoint is done by calling `NetApp::endpoint`.
+/// An endpoint is identified primarily by its path, which is specified
+/// at creation time.
+///
+/// An `Endpoint` is used both to send requests to remote nodes,
+/// and to specify the handler for such requests on the local node.
+/// The type `H` represents the type of the handler object for
+/// endpoint messages (see `StreamingEndpointHandler`).
+pub struct Endpoint<M, H>
+where
+	M: Message,
+	H: StreamingEndpointHandler<M>,
+{
+	_phantom: PhantomData<M>,
+	netapp: Arc<NetApp>,
+	path: String,
+	handler: ArcSwapOption<H>,
+}
+
+impl<M, H> Endpoint<M, H>
+where
+	M: Message,
+	H: StreamingEndpointHandler<M>,
+{
+	pub(crate) fn new(netapp: Arc<NetApp>, path: String) -> Self {
+		Self {
+			_phantom: PhantomData::default(),
+			netapp,
+			path,
+			handler: ArcSwapOption::from(None),
+		}
+	}
+
+	/// Get the path of this endpoint
+	pub fn path(&self) -> &str {
+		&self.path
+	}
+
+	/// Set the object that is responsible of handling requests to
+	/// this endpoint on the local node.
+	pub fn set_handler(&self, h: Arc<H>) {
+		self.handler.swap(Some(h));
+	}
+
+	/// Call this endpoint on a remote node (or on the local node,
+	/// for that matter). This function invokes the full version that
+	/// allows to attach a stream to the request and to
+	/// receive such a stream attached to the response.
+	pub async fn call_streaming<T>(
+		&self,
+		target: &NodeID,
+		req: T,
+		prio: RequestPriority,
+	) -> Result<Resp<M>, Error>
+	where
+		T: IntoReq<M>,
+	{
+		if *target == self.netapp.id {
+			match self.handler.load_full() {
+				None => Err(Error::NoHandler),
+				Some(h) => Ok(h.handle(req.into_req_local(), self.netapp.id).await),
+			}
+		} else {
+			let conn = self
+				.netapp
+				.client_conns
+				.read()
+				.unwrap()
+				.get(target)
+				.cloned();
+			match conn {
+				None => Err(Error::Message(format!(
+					"Not connected: {}",
+					hex::encode(&target[..8])
+				))),
+				Some(c) => c.call(req.into_req()?, self.path.as_str(), prio).await,
+			}
+		}
+	}
+
+	/// Call this endpoint on a remote node. This function is the simplified
+	/// version that doesn't allow to have streams attached to the request
+	/// or the response; see `call_streaming` for the full version.
+	pub async fn call(
+		&self,
+		target: &NodeID,
+		req: M,
+		prio: RequestPriority,
+	) -> Result<<M as Message>::Response, Error> {
+		Ok(self.call_streaming(target, req, prio).await?.into_msg())
+	}
+}
+
+// ---- Internal stuff ----
+
+pub(crate) type DynEndpoint = Box<dyn GenericEndpoint + Send + Sync>;
+
+#[async_trait]
+pub(crate) trait GenericEndpoint {
+	async fn handle(&self, req_enc: ReqEnc, from: NodeID) -> Result<RespEnc, Error>;
+	fn drop_handler(&self);
+	fn clone_endpoint(&self) -> DynEndpoint;
+}
+
+#[derive(Clone)]
+pub(crate) struct EndpointArc<M, H>(pub(crate) Arc<Endpoint<M, H>>)
+where
+	M: Message,
+	H: StreamingEndpointHandler<M>;
+
+#[async_trait]
+impl<M, H> GenericEndpoint for EndpointArc<M, H>
+where
+	M: Message,
+	H: StreamingEndpointHandler<M> + 'static,
+{
+	async fn handle(&self, req_enc: ReqEnc, from: NodeID) -> Result<RespEnc, Error> {
+		match self.0.handler.load_full() {
+			None => Err(Error::NoHandler),
+			Some(h) => {
+				let req = Req::from_enc(req_enc)?;
+				let res = h.handle(req, from).await;
+				Ok(res.into_enc()?)
+			}
+		}
+	}
+
+	fn drop_handler(&self) {
+		self.0.handler.swap(None);
+	}
+
+	fn clone_endpoint(&self) -> DynEndpoint {
+		Box::new(Self(self.0.clone()))
+	}
+}

src/error.rs

@@ -1,6 +1,6 @@
-use err_derive::Error;
 use std::io;
 
+use err_derive::Error;
 use log::error;
 
 #[derive(Debug, Error)]
@@ -22,11 +22,29 @@ pub enum Error {
 	#[error(display = "Handshake error: {}", _0)]
 	Handshake(#[error(source)] kuska_handshake::async_std::Error),
 
+	#[error(display = "UTF8 error: {}", _0)]
+	UTF8(#[error(source)] std::string::FromUtf8Error),
+
+	#[error(display = "Framing protocol error")]
+	Framing,
+
+	#[error(display = "Remote error ({:?}): {}", _0, _1)]
+	Remote(io::ErrorKind, String),
+
+	#[error(display = "Request ID collision")]
+	IdCollision,
+
 	#[error(display = "{}", _0)]
 	Message(String),
 
-	#[error(display = "Remote error: {}", _0)]
-	Remote(String),
+	#[error(display = "No handler / shutting down")]
+	NoHandler,
+
+	#[error(display = "Connection closed")]
+	ConnectionClosed,
+
+	#[error(display = "Version mismatch: {}", _0)]
+	VersionMismatch(String),
 }
 
 impl<T> From<tokio::sync::watch::error::SendError<T>> for Error {
@@ -57,3 +75,52 @@ where
 		};
 	}
 }
+
+impl<E, T> LogError for Result<T, E>
+where
+	T: LogError,
+	E: Into<Error>,
+{
+	fn log_err(self, msg: &'static str) {
+		match self {
+			Err(e) => error!("Error: {}: {}", msg, Into::<Error>::into(e)),
+			Ok(x) => x.log_err(msg),
+		}
+	}
+}
+
+// ---- Helpers for serializing I/O Errors
+
+pub(crate) fn u8_to_io_errorkind(v: u8) -> std::io::ErrorKind {
+	use std::io::ErrorKind;
+	match v {
+		101 => ErrorKind::ConnectionAborted,
+		102 => ErrorKind::BrokenPipe,
+		103 => ErrorKind::WouldBlock,
+		104 => ErrorKind::InvalidInput,
+		105 => ErrorKind::InvalidData,
+		106 => ErrorKind::TimedOut,
+		107 => ErrorKind::Interrupted,
+		108 => ErrorKind::UnexpectedEof,
+		109 => ErrorKind::OutOfMemory,
+		110 => ErrorKind::ConnectionReset,
+		_ => ErrorKind::Other,
+	}
+}
+
+pub(crate) fn io_errorkind_to_u8(kind: std::io::ErrorKind) -> u8 {
+	use std::io::ErrorKind;
+	match kind {
+		ErrorKind::ConnectionAborted => 101,
+		ErrorKind::BrokenPipe => 102,
+		ErrorKind::WouldBlock => 103,
+		ErrorKind::InvalidInput => 104,
+		ErrorKind::InvalidData => 105,
+		ErrorKind::TimedOut => 106,
+		ErrorKind::Interrupted => 107,
+		ErrorKind::UnexpectedEof => 108,
+		ErrorKind::OutOfMemory => 109,
+		ErrorKind::ConnectionReset => 110,
+		_ => 100,
+	}
+}

src/lib.rs

@@ -13,18 +13,23 @@
 //! about message priorization.
 //! Also check out the examples to learn how to use this crate.
 
-#![feature(map_first_last)]
-
+pub mod bytes_buf;
 pub mod error;
+pub mod stream;
 pub mod util;
 
+pub mod endpoint;
 pub mod message;
-pub mod proto;
 
-mod conn;
+mod client;
+mod recv;
+mod send;
+mod server;
 
 pub mod netapp;
 pub mod peering;
 
-pub use netapp::*;
-pub use util::NodeID;
+pub use crate::netapp::*;
+
+#[cfg(test)]
+mod test;

src/message.rs

@@ -1,36 +1,522 @@
-use std::net::IpAddr;
+use std::fmt;
+use std::marker::PhantomData;
+use std::sync::Arc;
 
+use bytes::{BufMut, Bytes, BytesMut};
+use rand::prelude::*;
 use serde::{Deserialize, Serialize};
 
-pub type MessageKind = u32;
+use futures::stream::StreamExt;
+
+use crate::error::*;
+use crate::stream::*;
+use crate::util::*;
+
+/// Priority of a request (click to read more about priorities).
+///
+/// This priority value is used to priorize messages
+/// in the send queue of the client, and their responses in the send queue of the
+/// server. Lower values mean higher priority.
+///
+/// This mechanism is usefull for messages bigger than the maximum chunk size
+/// (set at `0x4000` bytes), such as large file transfers.
+/// In such case, all of the messages in the send queue with the highest priority
+/// will take turns to send individual chunks, in a round-robin fashion.
+/// Once all highest priority messages are sent successfully, the messages with
+/// the next highest priority will begin being sent in the same way.
+///
+/// The same priority value is given to a request and to its associated response.
+pub type RequestPriority = u8;
+
+/// Priority class: high
+pub const PRIO_HIGH: RequestPriority = 0x20;
+/// Priority class: normal
+pub const PRIO_NORMAL: RequestPriority = 0x40;
+/// Priority class: background
+pub const PRIO_BACKGROUND: RequestPriority = 0x80;
+/// Priority: primary among given class
+pub const PRIO_PRIMARY: RequestPriority = 0x00;
+/// Priority: secondary among given class (ex: `PRIO_HIGH | PRIO_SECONDARY`)
+pub const PRIO_SECONDARY: RequestPriority = 0x01;
+
+// ----
+
+/// An order tag can be added to a message or a response to indicate
+/// whether it should be sent after or before other messages with order tags
+/// referencing a same stream
+#[derive(Clone, Copy, Serialize, Deserialize, Debug)]
+pub struct OrderTag(pub(crate) u64, pub(crate) u64);
+
+/// A stream is an opaque identifier that defines a set of messages
+/// or responses that are ordered wrt one another using to order tags.
+#[derive(Clone, Copy)]
+pub struct OrderTagStream(u64);
+
+impl OrderTag {
+	/// Create a new stream from which to generate order tags. Example:
+	/// ```ignore
+	/// let stream = OrderTag.stream();
+	/// let tag_1 = stream.order(1);
+	/// let tag_2 = stream.order(2);
+	/// ```
+	pub fn stream() -> OrderTagStream {
+		OrderTagStream(thread_rng().gen())
+	}
+}
+impl OrderTagStream {
+	/// Create the order tag for message `order` in this stream
+	pub fn order(&self, order: u64) -> OrderTag {
+		OrderTag(self.0, order)
+	}
+}
+
+// ----
 
 /// This trait should be implemented by all messages your application
-/// wants to handle (click to read more).
-///
-/// It defines a `KIND`, which should be a **unique**
-/// `u32` that distinguishes these messages from other types of messages
-/// (it is used by our communication protocol), as well as an associated
-/// `Response` type that defines the type of the response that is given
-/// to the message. It is your responsibility to ensure that `KIND` is a
-/// unique `u32` that is not used by any other protocol messages.
-/// All `KIND` values of the form `0x42xxxxxx` are reserved by the netapp
-/// crate for internal purposes.
-///
-/// A handler for this message has type `Self -> Self::Response`.
-/// If you need to return an error, the `Response` type should be
-/// a `Result<_, _>`.
-pub trait Message: Serialize + for<'de> Deserialize<'de> + Send + Sync {
-	const KIND: MessageKind;
-	type Response: Serialize + for<'de> Deserialize<'de> + Send + Sync;
+/// wants to handle. It specifies which data type should be sent
+/// as a response to this message in the RPC protocol.
+pub trait Message: Serialize + for<'de> Deserialize<'de> + Send + Sync + 'static {
+	/// The type of the response that is sent in response to this message
+	type Response: Serialize + for<'de> Deserialize<'de> + Send + Sync + 'static;
 }
 
-#[derive(Serialize, Deserialize)]
-pub(crate) struct HelloMessage {
-	pub server_addr: Option<IpAddr>,
-	pub server_port: u16,
+// ----
+
+/// The Req<M> is a helper object used to create requests and attach them
+/// a stream of data. If the stream is a fixed Bytes and not a ByteStream,
+/// Req<M> is cheaply clonable to allow the request to be sent to different
+/// peers (Clone will panic if the stream is a ByteStream).
+pub struct Req<M: Message> {
+	pub(crate) msg: Arc<M>,
+	pub(crate) msg_ser: Option<Bytes>,
+	pub(crate) stream: AttachedStream,
+	pub(crate) order_tag: Option<OrderTag>,
 }
 
-impl Message for HelloMessage {
-	const KIND: MessageKind = 0x42000001;
-	type Response = ();
+impl<M: Message> Req<M> {
+	/// Creates a new request from a base message `M`
+	pub fn new(v: M) -> Result<Self, Error> {
+		Ok(v.into_req()?)
+	}
+
+	/// Attach a stream to message in request, where the stream is streamed
+	/// from a fixed `Bytes` buffer
+	pub fn with_stream_from_buffer(self, b: Bytes) -> Self {
+		Self {
+			stream: AttachedStream::Fixed(b),
+			..self
+		}
+	}
+
+	/// Attach a stream to message in request, where the stream is
+	/// an instance of `ByteStream`. Note than when a `Req<M>` has an attached
+	/// stream which is a `ByteStream` instance, it can no longer be cloned
+	/// to be sent to different nodes (`.clone()` will panic)
+	pub fn with_stream(self, b: ByteStream) -> Self {
+		Self {
+			stream: AttachedStream::Stream(b),
+			..self
+		}
+	}
+
+	/// Add an order tag to this request to indicate in which order it should
+	/// be sent.
+	pub fn with_order_tag(self, order_tag: OrderTag) -> Self {
+		Self {
+			order_tag: Some(order_tag),
+			..self
+		}
+	}
+
+	/// Get a reference to the message `M` contained in this request
+	pub fn msg(&self) -> &M {
+		&self.msg
+	}
+
+	/// Takes out the stream attached to this request, if any
+	pub fn take_stream(&mut self) -> Option<ByteStream> {
+		std::mem::replace(&mut self.stream, AttachedStream::None).into_stream()
+	}
+
+	pub(crate) fn into_enc(
+		self,
+		prio: RequestPriority,
+		path: Bytes,
+		telemetry_id: Bytes,
+	) -> ReqEnc {
+		ReqEnc {
+			prio,
+			path,
+			telemetry_id,
+			msg: self.msg_ser.unwrap(),
+			stream: self.stream.into_stream(),
+			order_tag: self.order_tag,
+		}
+	}
+
+	pub(crate) fn from_enc(enc: ReqEnc) -> Result<Self, rmp_serde::decode::Error> {
+		let msg = rmp_serde::decode::from_read_ref(&enc.msg)?;
+		Ok(Req {
+			msg: Arc::new(msg),
+			msg_ser: Some(enc.msg),
+			stream: enc
+				.stream
+				.map(AttachedStream::Stream)
+				.unwrap_or(AttachedStream::None),
+			order_tag: enc.order_tag,
+		})
+	}
+}
+
+/// `IntoReq<M>` represents any object that can be transformed into `Req<M>`
+pub trait IntoReq<M: Message> {
+	/// Transform the object into a `Req<M>`, serializing the message M
+	/// to be sent to remote nodes
+	fn into_req(self) -> Result<Req<M>, rmp_serde::encode::Error>;
+	/// Transform the object into a `Req<M>`, skipping the serialization
+	/// of message M, in the case we are not sending this RPC message to
+	/// a remote node
+	fn into_req_local(self) -> Req<M>;
+}
+
+impl<M: Message> IntoReq<M> for M {
+	fn into_req(self) -> Result<Req<M>, rmp_serde::encode::Error> {
+		let msg_ser = rmp_to_vec_all_named(&self)?;
+		Ok(Req {
+			msg: Arc::new(self),
+			msg_ser: Some(Bytes::from(msg_ser)),
+			stream: AttachedStream::None,
+			order_tag: None,
+		})
+	}
+	fn into_req_local(self) -> Req<M> {
+		Req {
+			msg: Arc::new(self),
+			msg_ser: None,
+			stream: AttachedStream::None,
+			order_tag: None,
+		}
+	}
+}
+
+impl<M: Message> IntoReq<M> for Req<M> {
+	fn into_req(self) -> Result<Req<M>, rmp_serde::encode::Error> {
+		Ok(self)
+	}
+	fn into_req_local(self) -> Req<M> {
+		self
+	}
+}
+
+impl<M: Message> Clone for Req<M> {
+	fn clone(&self) -> Self {
+		let stream = match &self.stream {
+			AttachedStream::None => AttachedStream::None,
+			AttachedStream::Fixed(b) => AttachedStream::Fixed(b.clone()),
+			AttachedStream::Stream(_) => {
+				panic!("Cannot clone a Req<_> with a non-buffer attached stream")
+			}
+		};
+		Self {
+			msg: self.msg.clone(),
+			msg_ser: self.msg_ser.clone(),
+			stream,
+			order_tag: self.order_tag,
+		}
+	}
+}
+
+impl<M> fmt::Debug for Req<M>
+where
+	M: Message + fmt::Debug,
+{
+	fn fmt(&self, f: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
+		write!(f, "Req[{:?}", self.msg)?;
+		match &self.stream {
+			AttachedStream::None => write!(f, "]"),
+			AttachedStream::Fixed(b) => write!(f, "; stream=buf:{}]", b.len()),
+			AttachedStream::Stream(_) => write!(f, "; stream]"),
+		}
+	}
+}
+
+// ----
+
+/// The Resp<M> represents a full response from a RPC that may have
+/// an attached stream.
+pub struct Resp<M: Message> {
+	pub(crate) _phantom: PhantomData<M>,
+	pub(crate) msg: M::Response,
+	pub(crate) stream: AttachedStream,
+	pub(crate) order_tag: Option<OrderTag>,
+}
+
+impl<M: Message> Resp<M> {
+	/// Creates a new response from a base response message
+	pub fn new(v: M::Response) -> Self {
+		Resp {
+			_phantom: Default::default(),
+			msg: v,
+			stream: AttachedStream::None,
+			order_tag: None,
+		}
+	}
+
+	/// Attach a stream to message in response, where the stream is streamed
+	/// from a fixed `Bytes` buffer
+	pub fn with_stream_from_buffer(self, b: Bytes) -> Self {
+		Self {
+			stream: AttachedStream::Fixed(b),
+			..self
+		}
+	}
+
+	/// Attach a stream to message in response, where the stream is
+	/// an instance of `ByteStream`.
+	pub fn with_stream(self, b: ByteStream) -> Self {
+		Self {
+			stream: AttachedStream::Stream(b),
+			..self
+		}
+	}
+
+	/// Add an order tag to this response to indicate in which order it should
+	/// be sent.
+	pub fn with_order_tag(self, order_tag: OrderTag) -> Self {
+		Self {
+			order_tag: Some(order_tag),
+			..self
+		}
+	}
+
+	/// Get a reference to the response message contained in this request
+	pub fn msg(&self) -> &M::Response {
+		&self.msg
+	}
+
+	/// Transforms the `Resp<M>` into the response message it contains,
+	/// dropping everything else (including attached data stream)
+	pub fn into_msg(self) -> M::Response {
+		self.msg
+	}
+
+	/// Transforms the `Resp<M>` into, on the one side, the response message
+	/// it contains, and on the other side, the associated data stream
+	/// if it exists
+	pub fn into_parts(self) -> (M::Response, Option<ByteStream>) {
+		(self.msg, self.stream.into_stream())
+	}
+
+	pub(crate) fn into_enc(self) -> Result<RespEnc, rmp_serde::encode::Error> {
+		Ok(RespEnc {
+			msg: rmp_to_vec_all_named(&self.msg)?.into(),
+			stream: self.stream.into_stream(),
+			order_tag: self.order_tag,
+		})
+	}
+
+	pub(crate) fn from_enc(enc: RespEnc) -> Result<Self, Error> {
+		let msg = rmp_serde::decode::from_read_ref(&enc.msg)?;
+		Ok(Self {
+			_phantom: Default::default(),
+			msg,
+			stream: enc
+				.stream
+				.map(AttachedStream::Stream)
+				.unwrap_or(AttachedStream::None),
+			order_tag: enc.order_tag,
+		})
+	}
+}
+
+impl<M> fmt::Debug for Resp<M>
+where
+	M: Message,
+	<M as Message>::Response: fmt::Debug,
+{
+	fn fmt(&self, f: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
+		write!(f, "Resp[{:?}", self.msg)?;
+		match &self.stream {
+			AttachedStream::None => write!(f, "]"),
+			AttachedStream::Fixed(b) => write!(f, "; stream=buf:{}]", b.len()),
+			AttachedStream::Stream(_) => write!(f, "; stream]"),
+		}
+	}
+}
+
+// ----
+
+pub(crate) enum AttachedStream {
+	None,
+	Fixed(Bytes),
+	Stream(ByteStream),
+}
+
+impl AttachedStream {
+	pub fn into_stream(self) -> Option<ByteStream> {
+		match self {
+			AttachedStream::None => None,
+			AttachedStream::Fixed(b) => Some(Box::pin(futures::stream::once(async move { Ok(b) }))),
+			AttachedStream::Stream(s) => Some(s),
+		}
+	}
+}
+
+// ---- ----
+
+/// Encoding for requests into a ByteStream:
+/// - priority: u8
+/// - path length: u8
+/// - path: [u8; path length]
+/// - telemetry id length: u8
+/// - telemetry id: [u8; telemetry id length]
+/// - msg len: u32
+/// - msg [u8; ..]
+/// - the attached stream as the rest of the encoded stream
+pub(crate) struct ReqEnc {
+	pub(crate) prio: RequestPriority,
+	pub(crate) path: Bytes,
+	pub(crate) telemetry_id: Bytes,
+	pub(crate) msg: Bytes,
+	pub(crate) stream: Option<ByteStream>,
+	pub(crate) order_tag: Option<OrderTag>,
+}
+
+impl ReqEnc {
+	pub(crate) fn encode(self) -> (ByteStream, Option<OrderTag>) {
+		let mut buf = BytesMut::with_capacity(
+			self.path.len() + self.telemetry_id.len() + self.msg.len() + 16,
+		);
+
+		buf.put_u8(self.prio);
+
+		buf.put_u8(self.path.len() as u8);
+		buf.put(self.path);
+
+		buf.put_u8(self.telemetry_id.len() as u8);
+		buf.put(&self.telemetry_id[..]);
+
+		buf.put_u32(self.msg.len() as u32);
+
+		let header = buf.freeze();
+
+		let res_stream: ByteStream = if let Some(stream) = self.stream {
+			Box::pin(futures::stream::iter([Ok(header), Ok(self.msg)]).chain(stream))
+		} else {
+			Box::pin(futures::stream::iter([Ok(header), Ok(self.msg)]))
+		};
+		(res_stream, self.order_tag)
+	}
+
+	pub(crate) async fn decode(stream: ByteStream) -> Result<Self, Error> {
+		Self::decode_aux(stream)
+			.await
+			.map_err(read_exact_error_to_error)
+	}
+
+	async fn decode_aux(stream: ByteStream) -> Result<Self, ReadExactError> {
+		let mut reader = ByteStreamReader::new(stream);
+
+		let prio = reader.read_u8().await?;
+
+		let path_len = reader.read_u8().await?;
+		let path = reader.read_exact(path_len as usize).await?;
+
+		let telemetry_id_len = reader.read_u8().await?;
+		let telemetry_id = reader.read_exact(telemetry_id_len as usize).await?;
+
+		let msg_len = reader.read_u32().await?;
+		let msg = reader.read_exact(msg_len as usize).await?;
+
+		Ok(Self {
+			prio,
+			path,
+			telemetry_id,
+			msg,
+			stream: Some(reader.into_stream()),
+			order_tag: None,
+		})
+	}
+}
+
+/// Encoding for responses into a ByteStream:
+/// IF SUCCESS:
+/// - 0: u8
+/// - msg len: u32
+/// - msg [u8; ..]
+/// - the attached stream as the rest of the encoded stream
+/// IF ERROR:
+/// - message length + 1: u8
+/// - error code: u8
+/// - message: [u8; message_length]
+pub(crate) struct RespEnc {
+	msg: Bytes,
+	stream: Option<ByteStream>,
+	order_tag: Option<OrderTag>,
+}
+
+impl RespEnc {
+	pub(crate) fn encode(resp: Result<Self, Error>) -> (ByteStream, Option<OrderTag>) {
+		match resp {
+			Ok(Self {
+				msg,
+				stream,
+				order_tag,
+			}) => {
+				let mut buf = BytesMut::with_capacity(4);
+				buf.put_u32(msg.len() as u32);
+				let header = buf.freeze();
+
+				let res_stream: ByteStream = if let Some(stream) = stream {
+					Box::pin(futures::stream::iter([Ok(header), Ok(msg)]).chain(stream))
+				} else {
+					Box::pin(futures::stream::iter([Ok(header), Ok(msg)]))
+				};
+				(res_stream, order_tag)
+			}
+			Err(err) => {
+				let err = std::io::Error::new(
+					std::io::ErrorKind::Other,
+					format!("netapp error: {}", err),
+				);
+				(
+					Box::pin(futures::stream::once(async move { Err(err) })),
+					None,
+				)
+			}
+		}
+	}
+
+	pub(crate) async fn decode(stream: ByteStream) -> Result<Self, Error> {
+		Self::decode_aux(stream)
+			.await
+			.map_err(read_exact_error_to_error)
+	}
+
+	async fn decode_aux(stream: ByteStream) -> Result<Self, ReadExactError> {
+		let mut reader = ByteStreamReader::new(stream);
+
+		let msg_len = reader.read_u32().await?;
+		let msg = reader.read_exact(msg_len as usize).await?;
+
+		// Check whether the response stream still has data or not.
+		// If no more data is coming, this will defuse the request canceller.
+		// If we didn't do this, and the client doesn't try to read from the stream,
+		// the request canceller doesn't know that we read everything and
+		// sends a cancellation message to the server (which they don't care about).
+		reader.fill_buffer().await;
+
+		Ok(Self {
+			msg,
+			stream: Some(reader.into_stream()),
+			order_tag: None,
+		})
+	}
+}
+
+fn read_exact_error_to_error(e: ReadExactError) -> Error {
+	match e {
+		ReadExactError::Stream(err) => Error::Remote(err.kind(), err.to_string()),
+		ReadExactError::UnexpectedEos => Error::Framing,
+	}
 }

src/netapp.rs

@@ -1,49 +1,58 @@
-use std::any::Any;
 use std::collections::HashMap;
 use std::net::{IpAddr, SocketAddr};
-use std::pin::Pin;
 use std::sync::{Arc, RwLock};
-use std::time::Instant;
 
-use std::future::Future;
+use log::{debug, error, info, trace, warn};
 
-use log::{debug, info};
-
-use arc_swap::{ArcSwap, ArcSwapOption};
-use bytes::Bytes;
+use arc_swap::ArcSwapOption;
+use async_trait::async_trait;
 
+use serde::{Deserialize, Serialize};
 use sodiumoxide::crypto::auth;
 use sodiumoxide::crypto::sign::ed25519;
-use tokio::net::{TcpListener, TcpStream};
 
-use crate::conn::*;
+use futures::stream::futures_unordered::FuturesUnordered;
+use futures::stream::StreamExt;
+use tokio::net::{TcpListener, TcpStream};
+use tokio::select;
+use tokio::sync::{mpsc, watch};
+
+use crate::client::*;
+use crate::endpoint::*;
 use crate::error::*;
 use crate::message::*;
-use crate::proto::*;
-use crate::util::*;
+use crate::server::*;
 
-type DynMsg = Box<dyn Any + Send + Sync + 'static>;
+/// A node's identifier, which is also its public cryptographic key
+pub type NodeID = sodiumoxide::crypto::sign::ed25519::PublicKey;
+/// A node's secret key
+pub type NodeKey = sodiumoxide::crypto::sign::ed25519::SecretKey;
+/// A network key
+pub type NetworkKey = sodiumoxide::crypto::auth::Key;
+
+/// Tag which is exchanged between client and server upon connection establishment
+/// to check that they are running compatible versions of Netapp,
+/// composed of 8 bytes for Netapp version and 8 bytes for client version
+pub(crate) type VersionTag = [u8; 16];
+
+/// Value of the Netapp version used in the version tag
+pub(crate) const NETAPP_VERSION_TAG: u64 = 0x6e65746170700005; // netapp 0x0005
+
+#[derive(Serialize, Deserialize, Debug)]
+pub(crate) struct HelloMessage {
+	pub server_addr: Option<IpAddr>,
+	pub server_port: u16,
+}
+
+impl Message for HelloMessage {
+	type Response = ();
+}
 
 type OnConnectHandler = Box<dyn Fn(NodeID, SocketAddr, bool) + Send + Sync>;
 type OnDisconnectHandler = Box<dyn Fn(NodeID, bool) + Send + Sync>;
 
-pub(crate) type LocalHandler =
-	Box<dyn Fn(DynMsg) -> Pin<Box<dyn Future<Output = DynMsg> + Sync + Send>> + Sync + Send>;
-pub(crate) type NetHandler = Box<
-	dyn Fn(NodeID, Bytes) -> Pin<Box<dyn Future<Output = Vec<u8>> + Sync + Send>> + Sync + Send,
->;
-
-pub(crate) struct Handler {
-	pub(crate) local_handler: LocalHandler,
-	pub(crate) net_handler: NetHandler,
-}
-
 /// NetApp is the main class that handles incoming and outgoing connections.
 ///
-/// The `request()` method can be used to send a message to any peer to which we have
-/// an outgoing connection, or to ourself. On the server side, these messages are
-/// processed by the handlers that have been defined using `add_msg_handler()`.
-///
 /// NetApp can be used in a stand-alone fashion or together with a peering strategy.
 /// If using it alone, you will want to set `on_connect` and `on_disconnect` events
 /// in order to manage information about the current peer list.
@@ -53,6 +62,8 @@ pub(crate) struct Handler {
 pub struct NetApp {
 	listen_params: ArcSwapOption<ListenParams>,
 
+	/// Version tag, 8 bytes for netapp version, 8 bytes for app version
+	pub version_tag: VersionTag,
 	/// Network secret key
 	pub netid: auth::Key,
 	/// Our peer ID
@@ -60,10 +71,12 @@ pub struct NetApp {
 	/// Private key associated with our peer ID
 	pub privkey: ed25519::SecretKey,
 
-	server_conns: RwLock<HashMap<NodeID, Arc<ServerConn>>>,
-	client_conns: RwLock<HashMap<NodeID, Arc<ClientConn>>>,
+	pub(crate) server_conns: RwLock<HashMap<NodeID, Arc<ServerConn>>>,
+	pub(crate) client_conns: RwLock<HashMap<NodeID, Arc<ClientConn>>>,
+
+	pub(crate) endpoints: RwLock<HashMap<String, DynEndpoint>>,
+	hello_endpoint: ArcSwapOption<Endpoint<HelloMessage, NetApp>>,
 
-	pub(crate) msg_handlers: ArcSwap<HashMap<MessageKind, Arc<Handler>>>,
 	on_connected_handler: ArcSwapOption<OnConnectHandler>,
 	on_disconnected_handler: ArcSwapOption<OnDisconnectHandler>,
 }
@@ -73,69 +86,40 @@ struct ListenParams {
 	public_addr: Option<IpAddr>,
 }
 
-async fn net_handler_aux<M, F, R>(handler: Arc<F>, remote: NodeID, bytes: Bytes) -> Vec<u8>
-where
-	M: Message + 'static,
-	F: Fn(NodeID, M) -> R + Send + Sync + 'static,
-	R: Future<Output = <M as Message>::Response> + Send + Sync,
-{
-	debug!(
-		"Handling message of kind {:08x} from {}",
-		M::KIND,
-		hex::encode(remote)
-	);
-	let begin_time = Instant::now();
-	let res = match rmp_serde::decode::from_read_ref::<_, M>(&bytes[..]) {
-		Ok(msg) => Ok(handler(remote, msg).await),
-		Err(e) => Err(e.to_string()),
-	};
-	let end_time = Instant::now();
-	debug!(
-		"Request {:08x} from {} handled in {}msec",
-		M::KIND,
-		hex::encode(remote),
-		(end_time - begin_time).as_millis()
-	);
-	rmp_to_vec_all_named(&res).unwrap_or_default()
-}
-
-async fn local_handler_aux<M, F, R>(handler: Arc<F>, remote: NodeID, msg: DynMsg) -> DynMsg
-where
-	M: Message + 'static,
-	F: Fn(NodeID, M) -> R + Send + Sync + 'static,
-	R: Future<Output = <M as Message>::Response> + Send + Sync,
-{
-	debug!("Handling message of kind {:08x} from ourself", M::KIND);
-	let msg = (msg as Box<dyn Any + 'static>).downcast::<M>().unwrap();
-	let res = handler(remote, *msg).await;
-	Box::new(res)
-}
-
 impl NetApp {
 	/// Creates a new instance of NetApp, which can serve either as a full p2p node,
 	/// or just as a passive client. To upgrade to a full p2p node, spawn a listener
 	/// using `.listen()`
 	///
 	/// Our Peer ID is the public key associated to the secret key given here.
-	pub fn new(netid: auth::Key, privkey: ed25519::SecretKey) -> Arc<Self> {
+	pub fn new(app_version_tag: u64, netid: auth::Key, privkey: ed25519::SecretKey) -> Arc<Self> {
+		let mut version_tag = [0u8; 16];
+		version_tag[0..8].copy_from_slice(&u64::to_be_bytes(NETAPP_VERSION_TAG)[..]);
+		version_tag[8..16].copy_from_slice(&u64::to_be_bytes(app_version_tag)[..]);
+
 		let id = privkey.public_key();
 		let netapp = Arc::new(Self {
 			listen_params: ArcSwapOption::new(None),
+			version_tag,
 			netid,
 			id,
 			privkey,
 			server_conns: RwLock::new(HashMap::new()),
 			client_conns: RwLock::new(HashMap::new()),
-			msg_handlers: ArcSwap::new(Arc::new(HashMap::new())),
+			endpoints: RwLock::new(HashMap::new()),
+			hello_endpoint: ArcSwapOption::new(None),
 			on_connected_handler: ArcSwapOption::new(None),
 			on_disconnected_handler: ArcSwapOption::new(None),
 		});
 
-		let netapp2 = netapp.clone();
-		netapp.add_msg_handler::<HelloMessage, _, _>(move |from: NodeID, msg: HelloMessage| {
-			netapp2.handle_hello_message(from, msg);
-			async {}
-		});
+		netapp
+			.hello_endpoint
+			.swap(Some(netapp.endpoint("__netapp/netapp.rs/Hello".into())));
+		netapp
+			.hello_endpoint
+			.load_full()
+			.unwrap()
+			.set_handler(netapp.clone());
 
 		netapp
 	}
@@ -162,72 +146,134 @@ impl NetApp {
 			.store(Some(Arc::new(Box::new(handler))));
 	}
 
-	/// Add a handler for a certain message type. Note that only one handler
-	/// can be specified for each message type.
-	/// The handler is an asynchronous function, i.e. a function that returns
-	/// a future.
-	pub fn add_msg_handler<M, F, R>(&self, handler: F)
+	/// Create a new endpoint with path `path`,
+	/// that handles messages of type `M`.
+	/// `H` is the type of the object that should handle requests
+	/// to this endpoint on the local node. If you don't want
+	/// to handle request on the local node (e.g. if this node
+	/// is only a client in the network), define the type `H`
+	/// to be `()`.
+	/// This function will panic if the endpoint has already been
+	/// created.
+	pub fn endpoint<M, H>(self: &Arc<Self>, path: String) -> Arc<Endpoint<M, H>>
 	where
 		M: Message + 'static,
-		F: Fn(NodeID, M) -> R + Send + Sync + 'static,
-		R: Future<Output = <M as Message>::Response> + Send + Sync + 'static,
+		H: StreamingEndpointHandler<M> + 'static,
 	{
-		let handler = Arc::new(handler);
-
-		let handler2 = handler.clone();
-		let net_handler = Box::new(move |remote: NodeID, bytes: Bytes| {
-			let fun: Pin<Box<dyn Future<Output = Vec<u8>> + Sync + Send>> =
-				Box::pin(net_handler_aux(handler2.clone(), remote, bytes));
-			fun
-		});
-
-		let self_id = self.id;
-		let local_handler = Box::new(move |msg: DynMsg| {
-			let fun: Pin<Box<dyn Future<Output = DynMsg> + Sync + Send>> =
-				Box::pin(local_handler_aux(handler.clone(), self_id, msg));
-			fun
-		});
-
-		let funs = Arc::new(Handler {
-			net_handler,
-			local_handler,
-		});
-
-		let mut handlers = self.msg_handlers.load().as_ref().clone();
-		handlers.insert(M::KIND, funs);
-		self.msg_handlers.store(Arc::new(handlers));
+		let endpoint = Arc::new(Endpoint::<M, H>::new(self.clone(), path.clone()));
+		let endpoint_arc = EndpointArc(endpoint.clone());
+		if self
+			.endpoints
+			.write()
+			.unwrap()
+			.insert(path.clone(), Box::new(endpoint_arc))
+			.is_some()
+		{
+			panic!("Redefining endpoint: {}", path);
+		};
+		endpoint
 	}
 
 	/// Main listening process for our app. This future runs during the whole
 	/// run time of our application.
 	/// If this is not called, the NetApp instance remains a passive client.
-	pub async fn listen(self: Arc<Self>, listen_addr: SocketAddr, public_addr: Option<IpAddr>) {
+	pub async fn listen(
+		self: Arc<Self>,
+		listen_addr: SocketAddr,
+		public_addr: Option<IpAddr>,
+		mut must_exit: watch::Receiver<bool>,
+	) {
 		let listen_params = ListenParams {
 			listen_addr,
 			public_addr,
 		};
-		self.listen_params.store(Some(Arc::new(listen_params)));
+		if self
+			.listen_params
+			.swap(Some(Arc::new(listen_params)))
+			.is_some()
+		{
+			error!("Trying to listen on NetApp but we're already listening!");
+		}
 
 		let listener = TcpListener::bind(listen_addr).await.unwrap();
 		info!("Listening on {}", listen_addr);
 
+		let (conn_in, mut conn_out) = mpsc::unbounded_channel();
+		let connection_collector = tokio::spawn(async move {
+			let mut collection = FuturesUnordered::new();
 			loop {
-			// The second item contains the IP and port of the new connection.
-			let (socket, _) = listener.accept().await.unwrap();
+				if collection.is_empty() {
+					match conn_out.recv().await {
+						Some(f) => collection.push(f),
+						None => break,
+					}
+				} else {
+					select! {
+						new_fut = conn_out.recv() => {
+							match new_fut {
+								Some(f) => collection.push(f),
+								None => break,
+							}
+						}
+						result = collection.next() => {
+							trace!("Collected connection: {:?}", result);
+						}
+					}
+				}
+			}
+			debug!("Collecting last open server connections.");
+			while let Some(conn_res) = collection.next().await {
+				trace!("Collected connection: {:?}", conn_res);
+			}
+			debug!("No more server connections to collect");
+		});
+
+		while !*must_exit.borrow_and_update() {
+			let (socket, peer_addr) = select! {
+				sockres = listener.accept() => {
+					match sockres {
+						Ok(x) => x,
+						Err(e) => {
+							warn!("Error in listener.accept: {}", e);
+							continue;
+						}
+					}
+				},
+				_ = must_exit.changed() => continue,
+			};
+
 			info!(
 				"Incoming connection from {}, negotiating handshake...",
-				match socket.peer_addr() {
-					Ok(x) => format!("{}", x),
-					Err(e) => format!("<invalid addr: {}>", e),
-				}
+				peer_addr
 			);
 			let self2 = self.clone();
-			tokio::spawn(async move {
-				ServerConn::run(self2, socket)
+			let must_exit2 = must_exit.clone();
+			conn_in
+				.send(tokio::spawn(async move {
+					ServerConn::run(self2, socket, must_exit2)
 						.await
 						.log_err("ServerConn::run");
-			});
+				}))
+				.log_err("Failed to send connection to connection collector");
 		}
+
+		drop(conn_in);
+
+		connection_collector
+			.await
+			.log_err("Failed to await for connection collector");
+	}
+
+	/// Drop all endpoint handlers, as well as handlers for connection/disconnection
+	/// events. (This disables the peering strategy)
+	///
+	/// Use this when terminating to break reference cycles
+	pub fn drop_all_handlers(&self) {
+		for (_, endpoint) in self.endpoints.read().unwrap().iter() {
+			endpoint.drop_handler();
+		}
+		self.on_connected_handler.store(None);
+		self.on_disconnected_handler.store(None);
 	}
 
 	/// Attempt to connect to a peer, given by its ip:port and its public key.
@@ -267,7 +313,7 @@ impl NetApp {
 			if let Some(c) = conn {
 				debug!(
 					"Closing connection to {} ({})",
-					hex::encode(c.peer_id),
+					hex::encode(&c.peer_id[..8]),
 					c.remote_addr
 				);
 				c.close();
@@ -288,27 +334,17 @@ impl NetApp {
 		});
 	}
 
-	/// Close the incoming connection from a certain client to us,
-	/// if such a connection is currently open.
-	pub fn server_disconnect(self: &Arc<Self>, id: &NodeID) {
-		let conn = self.server_conns.read().unwrap().get(id).cloned();
-		if let Some(c) = conn {
-			debug!(
-				"Closing incoming connection from {} ({})",
-				hex::encode(c.peer_id),
-				c.remote_addr
-			);
-			c.close();
-		}
-	}
-
 	// Called from conn.rs when an incoming connection is successfully established
 	// Registers the connection in our list of connections
 	// Do not yet call the on_connected handler, because we don't know if the remote
 	// has an actual IP address and port we can call them back on.
 	// We will know this when they send a Hello message, which is handled below.
 	pub(crate) fn connected_as_server(&self, id: NodeID, conn: Arc<ServerConn>) {
-		info!("Accepted connection from {}", hex::encode(id));
+		info!(
+			"Accepted connection from {} at {}",
+			hex::encode(&id[..8]),
+			conn.remote_addr
+		);
 
 		self.server_conns.write().unwrap().insert(id, conn);
 	}
@@ -318,21 +354,12 @@ impl NetApp {
 	// At this point we know they are a full network member, and not just a client,
 	// and we call the on_connected handler so that the peering strategy knows
 	// we have a new potential peer
-	fn handle_hello_message(&self, id: NodeID, msg: HelloMessage) {
-		if let Some(h) = self.on_connected_handler.load().as_ref() {
-			if let Some(c) = self.server_conns.read().unwrap().get(&id) {
-				let remote_ip = msg.server_addr.unwrap_or_else(|| c.remote_addr.ip());
-				let remote_addr = SocketAddr::new(remote_ip, msg.server_port);
-				h(id, remote_addr, true);
-			}
-		}
-	}
 
 	// Called from conn.rs when an incoming connection is closed.
 	// We deregister the connection from server_conns and call the
 	// handler registered by on_disconnected
 	pub(crate) fn disconnected_as_server(&self, id: &NodeID, conn: Arc<ServerConn>) {
-		info!("Connection from {} closed", hex::encode(id));
+		info!("Connection from {} closed", hex::encode(&id[..8]));
 
 		let mut conn_list = self.server_conns.write().unwrap();
 		if let Some(c) = conn_list.get(id) {
@@ -355,7 +382,7 @@ impl NetApp {
 	// they know on which port to call us back. (TODO: don't do this if we are
 	// just a simple client and not a full p2p node)
 	pub(crate) fn connected_as_client(&self, id: NodeID, conn: Arc<ClientConn>) {
-		info!("Connection established to {}", hex::encode(id));
+		info!("Connection established to {}", hex::encode(&id[..8]));
 
 		{
 			let old_c_opt = self.client_conns.write().unwrap().insert(id, conn.clone());
@@ -371,8 +398,11 @@ impl NetApp {
 		if let Some(lp) = self.listen_params.load_full() {
 			let server_addr = lp.public_addr;
 			let server_port = lp.listen_addr.port();
+			let hello_endpoint = self.hello_endpoint.load_full().unwrap();
 			tokio::spawn(async move {
-				conn.request(
+				hello_endpoint
+					.call(
+						&conn.peer_id,
 						HelloMessage {
 							server_addr,
 							server_port,
@@ -380,6 +410,7 @@ impl NetApp {
 						PRIO_NORMAL,
 					)
 					.await
+					.map(|_| ())
 					.log_err("Sending hello message");
 			});
 		}
@@ -389,7 +420,7 @@ impl NetApp {
 	// The connection is removed from conn_list, and the on_disconnected handler
 	// is called.
 	pub(crate) fn disconnected_as_client(&self, id: &NodeID, conn: Arc<ClientConn>) {
-		info!("Connection to {} closed", hex::encode(id));
+		info!("Connection to {} closed", hex::encode(&id[..8]));
 		let mut conn_list = self.client_conns.write().unwrap();
 		if let Some(c) = conn_list.get(id) {
 			if Arc::ptr_eq(c, &conn) {
@@ -404,44 +435,17 @@ impl NetApp {
 		// else case: happens if connection was removed in .disconnect()
 		// in which case on_disconnected_handler was already called
 	}
+}
 
-	/// Send a message to a remote host to which a client connection is already
-	/// established, and await their response. The target is the id of the peer we
-	/// want to send the message to.
-	/// The priority is an `u8`, with lower numbers meaning highest priority.
-	pub async fn request<T>(
-		&self,
-		target: &NodeID,
-		rq: T,
-		prio: RequestPriority,
-	) -> Result<<T as Message>::Response, Error>
-	where
-		T: Message + 'static,
-	{
-		if *target == self.id {
-			let handler = self.msg_handlers.load().get(&T::KIND).cloned();
-			match handler {
-				None => Err(Error::Message(format!(
-					"No handler registered for message kind {:08x}",
-					T::KIND
-				))),
-				Some(h) => {
-					let local_handler = &h.local_handler;
-					let res = local_handler(Box::new(rq)).await;
-					let res_t = (res as Box<dyn Any + 'static>)
-						.downcast::<<T as Message>::Response>()
-						.unwrap();
-					Ok(*res_t)
-				}
-			}
-		} else {
-			let conn = self.client_conns.read().unwrap().get(target).cloned();
-			match conn {
-				None => Err(Error::Message(format!(
-					"Not connected: {}",
-					hex::encode(target)
-				))),
-				Some(c) => c.request(rq, prio).await,
+#[async_trait]
+impl EndpointHandler<HelloMessage> for NetApp {
+	async fn handle(self: &Arc<Self>, msg: &HelloMessage, from: NodeID) {
+		debug!("Hello from {:?}: {:?}", hex::encode(&from[..8]), msg);
+		if let Some(h) = self.on_connected_handler.load().as_ref() {
+			if let Some(c) = self.server_conns.read().unwrap().get(&from) {
+				let remote_ip = msg.server_addr.unwrap_or_else(|| c.remote_addr.ip());
+				let remote_addr = SocketAddr::new(remote_ip, msg.server_port);
+				h(from, remote_addr, true);
 			}
 		}
 	}

src/peering/basalt.rs

@@ -3,6 +3,7 @@ use std::net::SocketAddr;
 use std::sync::{Arc, RwLock};
 use std::time::Duration;
 
+use async_trait::async_trait;
 use log::{debug, info, trace, warn};
 use lru::LruCache;
 use rand::{thread_rng, Rng};
@@ -10,9 +11,11 @@ use serde::{Deserialize, Serialize};
 
 use sodiumoxide::crypto::hash;
 
+use tokio::sync::watch;
+
+use crate::endpoint::*;
 use crate::message::*;
 use crate::netapp::*;
-use crate::proto::*;
 use crate::NodeID;
 
 // -- Protocol messages --
@@ -21,7 +24,6 @@ use crate::NodeID;
 struct PullMessage {}
 
 impl Message for PullMessage {
-	const KIND: MessageKind = 0x42001100;
 	type Response = PushMessage;
 }
 
@@ -31,7 +33,6 @@ struct PushMessage {
 }
 
 impl Message for PushMessage {
-	const KIND: MessageKind = 0x42001101;
 	type Response = ();
 }
 
@@ -137,7 +138,7 @@ impl BasaltView {
 			let mut ret = vec![];
 			let mut rng = thread_rng();
 			for _i in 0..count {
-				let idx = rng.gen_range(0, possibles.len());
+				let idx = rng.gen_range(0..possibles.len());
 				ret.push(self.slots[possibles[idx]].peer.unwrap());
 			}
 			ret
@@ -236,6 +237,8 @@ pub struct BasaltParams {
 
 pub struct Basalt {
 	netapp: Arc<NetApp>,
+	pull_endpoint: Arc<Endpoint<PullMessage, Self>>,
+	push_endpoint: Arc<Endpoint<PushMessage, Self>>,
 
 	param: BasaltParams,
 	bootstrap_peers: Vec<Peer>,
@@ -264,6 +267,8 @@ impl Basalt {
 
 		let basalt = Arc::new(Self {
 			netapp: netapp.clone(),
+			pull_endpoint: netapp.endpoint("__netapp/peering/basalt.rs/Pull".into()),
+			push_endpoint: netapp.endpoint("__netapp/peering/basalt.rs/Push".into()),
 			param,
 			bootstrap_peers,
 			view: RwLock::new(view),
@@ -271,6 +276,9 @@ impl Basalt {
 			backlog: RwLock::new(backlog),
 		});
 
+		basalt.pull_endpoint.set_handler(basalt.clone());
+		basalt.push_endpoint.set_handler(basalt.clone());
+
 		let basalt2 = basalt.clone();
 		netapp.on_connected(move |id: NodeID, addr: SocketAddr, is_incoming: bool| {
 			basalt2.on_connected(id, addr, is_incoming);
@@ -281,18 +289,6 @@ impl Basalt {
 			basalt2.on_disconnected(id, is_incoming);
 		});
 
-		let basalt2 = basalt.clone();
-		netapp.add_msg_handler::<PullMessage, _, _>(move |_from: NodeID, _pullmsg: PullMessage| {
-			let push_msg = basalt2.make_push_message();
-			async move { push_msg }
-		});
-
-		let basalt2 = basalt.clone();
-		netapp.add_msg_handler::<PushMessage, _, _>(move |_from: NodeID, push_msg: PushMessage| {
-			basalt2.handle_peer_list(&push_msg.peers[..]);
-			async move {}
-		});
-
 		basalt
 	}
 
@@ -309,18 +305,19 @@ impl Basalt {
 			.collect::<Vec<_>>()
 	}
 
-	pub async fn run(self: Arc<Self>) {
+	pub async fn run(self: Arc<Self>, must_exit: watch::Receiver<bool>) {
 		for peer in self.bootstrap_peers.iter() {
 			tokio::spawn(self.clone().try_connect(*peer));
 		}
 
-		let pushpull_loop = self.clone().run_pushpull_loop();
-		let reset_loop = self.run_reset_loop();
-		tokio::join!(pushpull_loop, reset_loop);
+		tokio::join!(
+			self.clone().run_pushpull_loop(must_exit.clone()),
+			self.clone().run_reset_loop(must_exit.clone()),
+		);
 	}
 
-	async fn run_pushpull_loop(self: Arc<Self>) {
-		loop {
+	async fn run_pushpull_loop(self: Arc<Self>, must_exit: watch::Receiver<bool>) {
+		while !*must_exit.borrow() {
 			tokio::time::sleep(self.param.exchange_interval).await;
 
 			let peers = self.view.read().unwrap().sample(2);
@@ -333,8 +330,8 @@ impl Basalt {
 
 	async fn do_pull(self: Arc<Self>, peer: NodeID) {
 		match self
-			.netapp
-			.request(&peer, PullMessage {}, PRIO_NORMAL)
+			.pull_endpoint
+			.call(&peer, PullMessage {}, PRIO_NORMAL)
 			.await
 		{
 			Ok(resp) => {
@@ -349,7 +346,7 @@ impl Basalt {
 
 	async fn do_push(self: Arc<Self>, peer: NodeID) {
 		let push_msg = self.make_push_message();
-		match self.netapp.request(&peer, push_msg, PRIO_NORMAL).await {
+		match self.push_endpoint.call(&peer, push_msg, PRIO_NORMAL).await {
 			Ok(_) => {
 				trace!("KYEV PEXo {}", hex::encode(peer));
 			}
@@ -366,8 +363,8 @@ impl Basalt {
 		}
 	}
 
-	async fn run_reset_loop(self: Arc<Self>) {
-		loop {
+	async fn run_reset_loop(self: Arc<Self>, must_exit: watch::Receiver<bool>) {
+		while !*must_exit.borrow() {
 			tokio::time::sleep(self.param.reset_interval).await;
 
 			{
@@ -469,6 +466,20 @@ impl Basalt {
 	}
 }
 
+#[async_trait]
+impl EndpointHandler<PullMessage> for Basalt {
+	async fn handle(self: &Arc<Self>, _pullmsg: &PullMessage, _from: NodeID) -> PushMessage {
+		self.make_push_message()
+	}
+}
+
+#[async_trait]
+impl EndpointHandler<PushMessage> for Basalt {
+	async fn handle(self: &Arc<Self>, pushmsg: &PushMessage, _from: NodeID) {
+		self.handle_peer_list(&pushmsg.peers[..]);
+	}
+}
+
 fn rand_seed() -> Seed {
 	let mut seed = [0u8; 32];
 	sodiumoxide::randombytes::randombytes_into(&mut seed[..]);

src/peering/fullmesh.rs

@@ -4,20 +4,30 @@ use std::sync::atomic::{self, AtomicU64};
 use std::sync::{Arc, RwLock};
 use std::time::{Duration, Instant};
 
+use arc_swap::ArcSwap;
+use async_trait::async_trait;
 use log::{debug, info, trace, warn};
 use serde::{Deserialize, Serialize};
 
+use tokio::select;
+use tokio::sync::watch;
+
 use sodiumoxide::crypto::hash;
 
-use crate::message::*;
+use crate::endpoint::*;
+use crate::error::*;
 use crate::netapp::*;
-use crate::proto::*;
+
+use crate::message::*;
 use crate::NodeID;
 
 const CONN_RETRY_INTERVAL: Duration = Duration::from_secs(30);
 const CONN_MAX_RETRIES: usize = 10;
-const PING_INTERVAL: Duration = Duration::from_secs(10);
+const PING_INTERVAL: Duration = Duration::from_secs(15);
 const LOOP_DELAY: Duration = Duration::from_secs(1);
+const FAILED_PING_THRESHOLD: usize = 4;
+
+const DEFAULT_PING_TIMEOUT_MILLIS: u64 = 10_000;
 
 // -- Protocol messages --
 
@@ -28,7 +38,6 @@ struct PingMessage {
 }
 
 impl Message for PingMessage {
-	const KIND: MessageKind = 0x42001000;
 	type Response = PingMessage;
 }
 
@@ -38,55 +47,100 @@ struct PeerListMessage {
 }
 
 impl Message for PeerListMessage {
-	const KIND: MessageKind = 0x42001001;
 	type Response = PeerListMessage;
 }
 
 // -- Algorithm data structures --
 
 #[derive(Debug)]
-struct PeerInfo {
+struct PeerInfoInternal {
+	// addr is the currently connected address,
+	// or the last address we were connected to,
+	// or an arbitrary address some other peer gave us
 	addr: SocketAddr,
+	// all_addrs contains all of the addresses everyone gave us
+	all_addrs: Vec<SocketAddr>,
+
 	state: PeerConnState,
+	last_send_ping: Option<Instant>,
 	last_seen: Option<Instant>,
 	ping: VecDeque<Duration>,
+	failed_pings: usize,
 }
 
+impl PeerInfoInternal {
+	fn new(addr: SocketAddr, state: PeerConnState) -> Self {
+		Self {
+			addr,
+			all_addrs: vec![addr],
+			state,
+			last_send_ping: None,
+			last_seen: None,
+			ping: VecDeque::new(),
+			failed_pings: 0,
+		}
+	}
+}
+
+/// Information that the full mesh peering strategy can return about the peers it knows of
 #[derive(Copy, Clone, Debug)]
-pub struct PeerInfoPub {
+pub struct PeerInfo {
+	/// The node's identifier (its public key)
 	pub id: NodeID,
+	/// The node's network address
 	pub addr: SocketAddr,
+	/// The current status of our connection to this node
 	pub state: PeerConnState,
+	/// The last time at which the node was seen
 	pub last_seen: Option<Instant>,
+	/// The average ping to this node  on recent observations (if at least one ping value is known)
 	pub avg_ping: Option<Duration>,
+	/// The maximum observed ping to this node on recent observations (if at least one
+	/// ping value is known)
 	pub max_ping: Option<Duration>,
+	/// The median ping to this node on recent observations (if at least one ping value
+	/// is known)
 	pub med_ping: Option<Duration>,
 }
 
-// PeerConnState: possible states for our tentative connections to given peer
-// This module is only interested in recording connection info for outgoing
-// TCP connections
-#[derive(Copy, Clone, Debug, PartialEq)]
+impl PeerInfo {
+	/// Returns true if we can currently send requests to this peer
+	pub fn is_up(&self) -> bool {
+		self.state.is_up()
+	}
+}
+
+/// PeerConnState: possible states for our tentative connections to given peer
+/// This structure is only interested in recording connection info for outgoing
+/// TCP connections
+#[derive(Copy, Clone, Debug, PartialEq, Eq)]
 pub enum PeerConnState {
-	// This entry represents ourself
+	/// This entry represents ourself (the local node)
 	Ourself,
 
-	// We currently have a connection to this peer
+	/// We currently have a connection to this peer
 	Connected,
 
-	// Our next connection tentative (the nth, where n is the first value)
-	// will be at given Instant
+	/// Our next connection tentative (the nth, where n is the first value of the tuple)
+	/// will be at given Instant
 	Waiting(usize, Instant),
 
-	// A connection tentative is in progress
+	/// A connection tentative is in progress (the nth, where n is the value stored)
 	Trying(usize),
 
-	// We abandonned trying to connect to this peer (too many failed attempts)
+	/// We abandonned trying to connect to this peer (too many failed attempts)
 	Abandonned,
 }
 
+impl PeerConnState {
+	/// Returns true if we can currently send requests to this peer
+	pub fn is_up(&self) -> bool {
+		matches!(self, Self::Ourself | Self::Connected)
+	}
+}
+
 struct KnownHosts {
-	list: HashMap<NodeID, PeerInfo>,
+	list: HashMap<NodeID, PeerInfoInternal>,
 	hash: hash::Digest,
 }
 
@@ -99,7 +153,7 @@ impl KnownHosts {
 	fn update_hash(&mut self) {
 		self.hash = Self::calculate_hash(&self.list);
 	}
-	fn map_into_vec(input: &HashMap<NodeID, PeerInfo>) -> Vec<(NodeID, SocketAddr)> {
+	fn map_into_vec(input: &HashMap<NodeID, PeerInfoInternal>) -> Vec<(NodeID, SocketAddr)> {
 		let mut list = Vec::with_capacity(input.len());
 		for (id, peer) in input.iter() {
 			if peer.state == PeerConnState::Connected || peer.state == PeerConnState::Ourself {
@@ -108,96 +162,106 @@ impl KnownHosts {
 		}
 		list
 	}
-	fn calculate_hash(input: &HashMap<NodeID, PeerInfo>) -> hash::Digest {
+	fn calculate_hash(input: &HashMap<NodeID, PeerInfoInternal>) -> hash::Digest {
 		let mut list = Self::map_into_vec(input);
 		list.sort();
 		let mut hash_state = hash::State::new();
 		for (id, addr) in list {
 			hash_state.update(&id[..]);
-			hash_state.update(&format!("{}", addr).into_bytes()[..]);
+			hash_state.update(&format!("{}\n", addr).into_bytes()[..]);
 		}
 		hash_state.finalize()
 	}
 }
 
+/// A "Full Mesh" peering strategy is a peering strategy that tries
+/// to establish and maintain a direct connection with all of the
+/// known nodes in the network.
 pub struct FullMeshPeeringStrategy {
 	netapp: Arc<NetApp>,
 	known_hosts: RwLock<KnownHosts>,
+	public_peer_list: ArcSwap<Vec<PeerInfo>>,
+
 	next_ping_id: AtomicU64,
+	ping_endpoint: Arc<Endpoint<PingMessage, Self>>,
+	peer_list_endpoint: Arc<Endpoint<PeerListMessage, Self>>,
+
+	ping_timeout_millis: AtomicU64,
 }
 
 impl FullMeshPeeringStrategy {
-	pub fn new(netapp: Arc<NetApp>, bootstrap_list: Vec<(NodeID, SocketAddr)>) -> Arc<Self> {
+	/// Create a new Full Mesh peering strategy.
+	/// The strategy will not be run until `.run()` is called and awaited.
+	/// Once that happens, the peering strategy will try to connect
+	/// to all of the nodes specified in the bootstrap list.
+	pub fn new(
+		netapp: Arc<NetApp>,
+		bootstrap_list: Vec<(NodeID, SocketAddr)>,
+		our_addr: Option<SocketAddr>,
+	) -> Arc<Self> {
 		let mut known_hosts = KnownHosts::new();
 		for (id, addr) in bootstrap_list {
 			if id != netapp.id {
 				known_hosts.list.insert(
 					id,
-					PeerInfo {
-						addr,
-						state: PeerConnState::Waiting(0, Instant::now()),
-						last_seen: None,
-						ping: VecDeque::new(),
-					},
+					PeerInfoInternal::new(addr, PeerConnState::Waiting(0, Instant::now())),
 				);
 			}
 		}
 
+		if let Some(addr) = our_addr {
+			known_hosts.list.insert(
+				netapp.id,
+				PeerInfoInternal::new(addr, PeerConnState::Ourself),
+			);
+		}
+
 		let strat = Arc::new(Self {
 			netapp: netapp.clone(),
 			known_hosts: RwLock::new(known_hosts),
+			public_peer_list: ArcSwap::new(Arc::new(Vec::new())),
 			next_ping_id: AtomicU64::new(42),
+			ping_endpoint: netapp.endpoint("__netapp/peering/fullmesh.rs/Ping".into()),
+			peer_list_endpoint: netapp.endpoint("__netapp/peering/fullmesh.rs/PeerList".into()),
+			ping_timeout_millis: DEFAULT_PING_TIMEOUT_MILLIS.into(),
 		});
 
-		let strat2 = strat.clone();
-		netapp.add_msg_handler::<PingMessage, _, _>(move |from: NodeID, ping: PingMessage| {
-			let ping_resp = PingMessage {
-				id: ping.id,
-				peer_list_hash: strat2.known_hosts.read().unwrap().hash,
-			};
-			debug!("Ping from {}", hex::encode(&from));
-			async move { ping_resp }
-		});
+		strat.update_public_peer_list(&strat.known_hosts.read().unwrap());
 
-		let strat2 = strat.clone();
-		netapp.add_msg_handler::<PeerListMessage, _, _>(
-			move |_from: NodeID, peer_list: PeerListMessage| {
-				strat2.handle_peer_list(&peer_list.list[..]);
-				let peer_list = KnownHosts::map_into_vec(&strat2.known_hosts.read().unwrap().list);
-				let resp = PeerListMessage { list: peer_list };
-				async move { resp }
-			},
-		);
+		strat.ping_endpoint.set_handler(strat.clone());
+		strat.peer_list_endpoint.set_handler(strat.clone());
 
 		let strat2 = strat.clone();
 		netapp.on_connected(move |id: NodeID, addr: SocketAddr, is_incoming: bool| {
 			let strat2 = strat2.clone();
-			tokio::spawn(strat2.on_connected(id, addr, is_incoming));
+			strat2.on_connected(id, addr, is_incoming);
 		});
 
 		let strat2 = strat.clone();
 		netapp.on_disconnected(move |id: NodeID, is_incoming: bool| {
 			let strat2 = strat2.clone();
-			tokio::spawn(strat2.on_disconnected(id, is_incoming));
+			strat2.on_disconnected(id, is_incoming);
 		});
 
 		strat
 	}
 
-	pub async fn run(self: Arc<Self>) {
-		loop {
+	/// Run the full mesh peering strategy.
+	/// This future exits when the `must_exit` watch becomes true.
+	pub async fn run(self: Arc<Self>, must_exit: watch::Receiver<bool>) {
+		while !*must_exit.borrow() {
 			// 1. Read current state: get list of connected peers (ping them)
 			let (to_ping, to_retry) = {
 				let known_hosts = self.known_hosts.read().unwrap();
-				debug!("known_hosts: {} peers", known_hosts.list.len());
+				trace!("known_hosts: {} peers", known_hosts.list.len());
 
 				let mut to_ping = vec![];
 				let mut to_retry = vec![];
 				for (id, info) in known_hosts.list.iter() {
-					debug!("{}, {:?}", hex::encode(id), info);
+					trace!("{}, {:?}", hex::encode(&id[..8]), info);
 					match info.state {
 						PeerConnState::Connected => {
-							let must_ping = match info.last_seen {
+							let must_ping = match info.last_send_ping {
 								None => true,
 								Some(t) => Instant::now() - t > PING_INTERVAL,
 							};
@@ -217,10 +281,17 @@ impl FullMeshPeeringStrategy {
 			};
 
 			// 2. Dispatch ping to hosts
-			trace!("to_ping: {} peers", to_retry.len());
+			trace!("to_ping: {} peers", to_ping.len());
+			if !to_ping.is_empty() {
+				let mut known_hosts = self.known_hosts.write().unwrap();
+				for id in to_ping.iter() {
+					known_hosts.list.get_mut(id).unwrap().last_send_ping = Some(Instant::now());
+				}
+				drop(known_hosts);
 				for id in to_ping {
 					tokio::spawn(self.clone().ping(id));
 				}
+			}
 
 			// 3. Try reconnects
 			trace!("to_retry: {} peers", to_retry.len());
@@ -231,15 +302,27 @@ impl FullMeshPeeringStrategy {
 						if let PeerConnState::Waiting(i, _) = h.state {
 							info!(
 								"Retrying connection to {} at {} ({})",
-								hex::encode(&id),
-								h.addr,
+								hex::encode(&id[..8]),
+								h.all_addrs
+									.iter()
+									.map(|x| format!("{}", x))
+									.collect::<Vec<_>>()
+									.join(", "),
 								i + 1
 							);
 							h.state = PeerConnState::Trying(i);
-							tokio::spawn(self.clone().try_connect(id, h.addr));
+
+							let alternate_addrs = h
+								.all_addrs
+								.iter()
+								.filter(|x| **x != h.addr)
+								.cloned()
+								.collect::<Vec<_>>();
+							tokio::spawn(self.clone().try_connect(id, h.addr, alternate_addrs));
 						}
 					}
 				}
+				self.update_public_peer_list(&known_hosts);
 			}
 
 			// 4. Sleep before next loop iteration
@@ -247,125 +330,26 @@ impl FullMeshPeeringStrategy {
 		}
 	}
 
-	async fn ping(self: Arc<Self>, id: NodeID) {
-		let peer_list_hash = self.known_hosts.read().unwrap().hash;
-		let ping_id = self.next_ping_id.fetch_add(1u64, atomic::Ordering::Relaxed);
-		let ping_time = Instant::now();
-		let ping_msg = PingMessage {
-			id: ping_id,
-			peer_list_hash,
-		};
-
-		debug!(
-			"Sending ping {} to {} at {:?}",
-			ping_id,
-			hex::encode(id),
-			ping_time
-		);
-		match self.netapp.request(&id, ping_msg, PRIO_HIGH).await {
-			Err(e) => warn!("Error pinging {}: {}", hex::encode(id), e),
-			Ok(ping_resp) => {
-				let resp_time = Instant::now();
-				debug!(
-					"Got ping response from {} at {:?}",
-					hex::encode(id),
-					resp_time
-				);
-				{
-					let mut known_hosts = self.known_hosts.write().unwrap();
-					if let Some(host) = known_hosts.list.get_mut(&id) {
-						host.last_seen = Some(resp_time);
-						host.ping.push_back(resp_time - ping_time);
-						while host.ping.len() > 10 {
-							host.ping.pop_front();
-						}
-					}
-				}
-				if ping_resp.peer_list_hash != peer_list_hash {
-					self.exchange_peers(&id).await;
-				}
-			}
-		}
+	/// Returns a list of currently known peers in the network.
+	pub fn get_peer_list(&self) -> Arc<Vec<PeerInfo>> {
+		self.public_peer_list.load_full()
 	}
 
-	async fn exchange_peers(self: Arc<Self>, id: &NodeID) {
-		let peer_list = KnownHosts::map_into_vec(&self.known_hosts.read().unwrap().list);
-		let pex_message = PeerListMessage { list: peer_list };
-		match self.netapp.request(id, pex_message, PRIO_BACKGROUND).await {
-			Err(e) => warn!("Error doing peer exchange: {}", e),
-			Ok(resp) => {
-				self.handle_peer_list(&resp.list[..]);
-			}
-		}
+	/// Set the timeout for ping messages, in milliseconds
+	pub fn set_ping_timeout_millis(&self, timeout: u64) {
+		self.ping_timeout_millis
+			.store(timeout, atomic::Ordering::Relaxed);
 	}
 
-	fn handle_peer_list(&self, list: &[(NodeID, SocketAddr)]) {
-		let mut known_hosts = self.known_hosts.write().unwrap();
-		for (id, addr) in list.iter() {
-			if !known_hosts.list.contains_key(id) {
-				known_hosts.list.insert(*id, self.new_peer(id, *addr));
-			}
-		}
-	}
+	// -- internal stuff --
 
-	async fn try_connect(self: Arc<Self>, id: NodeID, addr: SocketAddr) {
-		let conn_result = self.netapp.clone().try_connect(addr, id).await;
-		if let Err(e) = conn_result {
-			warn!("Error connecting to {}: {}", hex::encode(id), e);
-			let mut known_hosts = self.known_hosts.write().unwrap();
-			if let Some(host) = known_hosts.list.get_mut(&id) {
-				host.state = match host.state {
-					PeerConnState::Trying(i) => {
-						if i >= CONN_MAX_RETRIES {
-							PeerConnState::Abandonned
-						} else {
-							PeerConnState::Waiting(i + 1, Instant::now() + CONN_RETRY_INTERVAL)
-						}
-					}
-					_ => PeerConnState::Waiting(0, Instant::now() + CONN_RETRY_INTERVAL),
-				};
-			}
-		}
-	}
-
-	async fn on_connected(self: Arc<Self>, id: NodeID, addr: SocketAddr, is_incoming: bool) {
-		if is_incoming {
-			if !self.known_hosts.read().unwrap().list.contains_key(&id) {
-				self.known_hosts
-					.write()
-					.unwrap()
-					.list
-					.insert(id, self.new_peer(&id, addr));
-			}
-		} else {
-			info!("Successfully connected to {} at {}", hex::encode(&id), addr);
-			let mut known_hosts = self.known_hosts.write().unwrap();
-			if let Some(host) = known_hosts.list.get_mut(&id) {
-				host.state = PeerConnState::Connected;
-				known_hosts.update_hash();
-			}
-		}
-	}
-
-	async fn on_disconnected(self: Arc<Self>, id: NodeID, is_incoming: bool) {
-		if !is_incoming {
-			info!("Connection to {} was closed", hex::encode(id));
-			let mut known_hosts = self.known_hosts.write().unwrap();
-			if let Some(host) = known_hosts.list.get_mut(&id) {
-				host.state = PeerConnState::Waiting(0, Instant::now());
-				known_hosts.update_hash();
-			}
-		}
-	}
-
-	pub fn get_peer_list(&self) -> Vec<PeerInfoPub> {
-		let known_hosts = self.known_hosts.read().unwrap();
-		let mut ret = Vec::with_capacity(known_hosts.list.len());
+	fn update_public_peer_list(&self, known_hosts: &KnownHosts) {
+		let mut pub_peer_list = Vec::with_capacity(known_hosts.list.len());
 		for (id, info) in known_hosts.list.iter() {
 			let mut pings = info.ping.iter().cloned().collect::<Vec<_>>();
 			pings.sort();
 			if !pings.is_empty() {
-				ret.push(PeerInfoPub {
+				pub_peer_list.push(PeerInfo {
 					id: *id,
 					addr: info.addr,
 					state: info.state,
@@ -380,7 +364,7 @@ impl FullMeshPeeringStrategy {
 					med_ping: Some(pings[pings.len() / 2]),
 				});
 			} else {
-				ret.push(PeerInfoPub {
+				pub_peer_list.push(PeerInfo {
 					id: *id,
 					addr: info.addr,
 					state: info.state,
@@ -391,20 +375,239 @@ impl FullMeshPeeringStrategy {
 				});
 			}
 		}
-		ret
+		self.public_peer_list.store(Arc::new(pub_peer_list));
 	}
 
-	fn new_peer(&self, id: &NodeID, addr: SocketAddr) -> PeerInfo {
+	async fn ping(self: Arc<Self>, id: NodeID) {
+		let peer_list_hash = self.known_hosts.read().unwrap().hash;
+		let ping_id = self.next_ping_id.fetch_add(1u64, atomic::Ordering::Relaxed);
+		let ping_time = Instant::now();
+		let ping_timeout =
+			Duration::from_millis(self.ping_timeout_millis.load(atomic::Ordering::Relaxed));
+		let ping_msg = PingMessage {
+			id: ping_id,
+			peer_list_hash,
+		};
+
+		debug!(
+			"Sending ping {} to {} at {:?}",
+			ping_id,
+			hex::encode(&id[..8]),
+			ping_time
+		);
+		let ping_response = select! {
+			r = self.ping_endpoint.call(&id, ping_msg, PRIO_HIGH) => r,
+			_ = tokio::time::sleep(ping_timeout) => Err(Error::Message("Ping timeout".into())),
+		};
+
+		match ping_response {
+			Err(e) => {
+				warn!("Error pinging {}: {}", hex::encode(&id[..8]), e);
+				let mut known_hosts = self.known_hosts.write().unwrap();
+				if let Some(host) = known_hosts.list.get_mut(&id) {
+					host.failed_pings += 1;
+					if host.failed_pings > FAILED_PING_THRESHOLD {
+						warn!(
+							"Too many failed pings from {}, closing connection.",
+							hex::encode(&id[..8])
+						);
+						// this will later update info in known_hosts
+						// through the disconnection handler
+						self.netapp.disconnect(&id);
+					}
+				}
+			}
+			Ok(ping_resp) => {
+				let resp_time = Instant::now();
+				debug!(
+					"Got ping response from {} at {:?}",
+					hex::encode(&id[..8]),
+					resp_time
+				);
+				{
+					let mut known_hosts = self.known_hosts.write().unwrap();
+					if let Some(host) = known_hosts.list.get_mut(&id) {
+						host.failed_pings = 0;
+						host.last_seen = Some(resp_time);
+						host.ping.push_back(resp_time - ping_time);
+						while host.ping.len() > 10 {
+							host.ping.pop_front();
+						}
+						self.update_public_peer_list(&known_hosts);
+					}
+				}
+				if ping_resp.peer_list_hash != peer_list_hash {
+					self.exchange_peers(&id).await;
+				}
+			}
+		}
+	}
+
+	async fn exchange_peers(self: Arc<Self>, id: &NodeID) {
+		let peer_list = KnownHosts::map_into_vec(&self.known_hosts.read().unwrap().list);
+		let pex_message = PeerListMessage { list: peer_list };
+		match self
+			.peer_list_endpoint
+			.call(id, pex_message, PRIO_BACKGROUND)
+			.await
+		{
+			Err(e) => warn!("Error doing peer exchange: {}", e),
+			Ok(resp) => {
+				self.handle_peer_list(&resp.list[..]);
+			}
+		}
+	}
+
+	fn handle_peer_list(&self, list: &[(NodeID, SocketAddr)]) {
+		let mut known_hosts = self.known_hosts.write().unwrap();
+
+		let mut changed = false;
+		for (id, addr) in list.iter() {
+			if let Some(kh) = known_hosts.list.get_mut(id) {
+				if !kh.all_addrs.contains(addr) {
+					kh.all_addrs.push(*addr);
+					changed = true;
+				}
+			} else {
+				known_hosts.list.insert(*id, self.new_peer(id, *addr));
+				changed = true;
+			}
+		}
+
+		if changed {
+			known_hosts.update_hash();
+			self.update_public_peer_list(&known_hosts);
+		}
+	}
+
+	async fn try_connect(
+		self: Arc<Self>,
+		id: NodeID,
+		default_addr: SocketAddr,
+		alternate_addrs: Vec<SocketAddr>,
+	) {
+		let conn_addr = {
+			let mut ret = None;
+			for addr in [default_addr].iter().chain(alternate_addrs.iter()) {
+				debug!("Trying address {} for peer {}", addr, hex::encode(&id[..8]));
+				match self.netapp.clone().try_connect(*addr, id).await {
+					Ok(()) => {
+						ret = Some(*addr);
+						break;
+					}
+					Err(e) => {
+						debug!(
+							"Error connecting to {} at {}: {}",
+							hex::encode(&id[..8]),
+							addr,
+							e
+						);
+					}
+				}
+			}
+			ret
+		};
+
+		if let Some(ok_addr) = conn_addr {
+			self.on_connected(id, ok_addr, false);
+		} else {
+			warn!(
+				"Could not connect to peer {} ({} addresses tried)",
+				hex::encode(&id[..8]),
+				1 + alternate_addrs.len()
+			);
+			let mut known_hosts = self.known_hosts.write().unwrap();
+			if let Some(host) = known_hosts.list.get_mut(&id) {
+				host.state = match host.state {
+					PeerConnState::Trying(i) => {
+						if i >= CONN_MAX_RETRIES {
+							PeerConnState::Abandonned
+						} else {
+							PeerConnState::Waiting(i + 1, Instant::now() + CONN_RETRY_INTERVAL)
+						}
+					}
+					_ => PeerConnState::Waiting(0, Instant::now() + CONN_RETRY_INTERVAL),
+				};
+				self.update_public_peer_list(&known_hosts);
+			}
+		}
+	}
+
+	fn on_connected(self: Arc<Self>, id: NodeID, addr: SocketAddr, is_incoming: bool) {
+		let mut known_hosts = self.known_hosts.write().unwrap();
+		if is_incoming {
+			if let Some(host) = known_hosts.list.get_mut(&id) {
+				if !host.all_addrs.contains(&addr) {
+					host.all_addrs.push(addr);
+				}
+			} else {
+				known_hosts.list.insert(id, self.new_peer(&id, addr));
+			}
+		} else {
+			info!(
+				"Successfully connected to {} at {}",
+				hex::encode(&id[..8]),
+				addr
+			);
+			if let Some(host) = known_hosts.list.get_mut(&id) {
+				host.state = PeerConnState::Connected;
+				host.addr = addr;
+				if !host.all_addrs.contains(&addr) {
+					host.all_addrs.push(addr);
+				}
+			} else {
+				known_hosts
+					.list
+					.insert(id, PeerInfoInternal::new(addr, PeerConnState::Connected));
+			}
+		}
+		known_hosts.update_hash();
+		self.update_public_peer_list(&known_hosts);
+	}
+
+	fn on_disconnected(self: Arc<Self>, id: NodeID, is_incoming: bool) {
+		if !is_incoming {
+			info!("Connection to {} was closed", hex::encode(&id[..8]));
+			let mut known_hosts = self.known_hosts.write().unwrap();
+			if let Some(host) = known_hosts.list.get_mut(&id) {
+				host.state = PeerConnState::Waiting(0, Instant::now());
+				known_hosts.update_hash();
+				self.update_public_peer_list(&known_hosts);
+			}
+		}
+	}
+
+	fn new_peer(&self, id: &NodeID, addr: SocketAddr) -> PeerInfoInternal {
 		let state = if *id == self.netapp.id {
 			PeerConnState::Ourself
 		} else {
 			PeerConnState::Waiting(0, Instant::now())
 		};
-		PeerInfo {
-			addr,
-			state,
-			last_seen: None,
-			ping: VecDeque::new(),
-		}
+		PeerInfoInternal::new(addr, state)
+	}
+}
+
+#[async_trait]
+impl EndpointHandler<PingMessage> for FullMeshPeeringStrategy {
+	async fn handle(self: &Arc<Self>, ping: &PingMessage, from: NodeID) -> PingMessage {
+		let ping_resp = PingMessage {
+			id: ping.id,
+			peer_list_hash: self.known_hosts.read().unwrap().hash,
+		};
+		debug!("Ping from {}", hex::encode(&from[..8]));
+		ping_resp
+	}
+}
+
+#[async_trait]
+impl EndpointHandler<PeerListMessage> for FullMeshPeeringStrategy {
+	async fn handle(
+		self: &Arc<Self>,
+		peer_list: &PeerListMessage,
+		_from: NodeID,
+	) -> PeerListMessage {
+		self.handle_peer_list(&peer_list.list[..]);
+		let peer_list = KnownHosts::map_into_vec(&self.known_hosts.read().unwrap().list);
+		PeerListMessage { list: peer_list }
 	}
 }

src/proto.rs

@@ -1,202 +0,0 @@
-use std::collections::{BTreeMap, HashMap, VecDeque};
-use std::sync::Arc;
-
-use log::trace;
-
-use futures::{AsyncReadExt, AsyncWriteExt};
-
-use tokio::sync::mpsc;
-
-use async_trait::async_trait;
-
-use crate::error::*;
-
-/// Priority of a request (click to read more about priorities).
-///
-/// This priority value is used to priorize messages
-/// in the send queue of the client, and their responses in the send queue of the
-/// server. Lower values mean higher priority.
-///
-/// This mechanism is usefull for messages bigger than the maximum chunk size
-/// (set at `0x4000` bytes), such as large file transfers.
-/// In such case, all of the messages in the send queue with the highest priority
-/// will take turns to send individual chunks, in a round-robin fashion.
-/// Once all highest priority messages are sent successfully, the messages with
-/// the next highest priority will begin being sent in the same way.
-///
-/// The same priority value is given to a request and to its associated response.
-pub type RequestPriority = u8;
-
-/// Priority class: high
-pub const PRIO_HIGH: RequestPriority = 0x20;
-/// Priority class: normal
-pub const PRIO_NORMAL: RequestPriority = 0x40;
-/// Priority class: background
-pub const PRIO_BACKGROUND: RequestPriority = 0x80;
-/// Priority: primary among given class
-pub const PRIO_PRIMARY: RequestPriority = 0x00;
-/// Priority: secondary among given class (ex: `PRIO_HIGH | PRIO_SECONDARY`)
-pub const PRIO_SECONDARY: RequestPriority = 0x01;
-
-const MAX_CHUNK_SIZE: usize = 0x4000;
-
-pub(crate) type RequestID = u16;
-
-struct SendQueueItem {
-	id: RequestID,
-	prio: RequestPriority,
-	data: Vec<u8>,
-	cursor: usize,
-}
-
-struct SendQueue {
-	items: BTreeMap<u8, VecDeque<SendQueueItem>>,
-}
-
-impl SendQueue {
-	fn new() -> Self {
-		Self {
-			items: BTreeMap::new(),
-		}
-	}
-	fn push(&mut self, item: SendQueueItem) {
-		let prio = item.prio;
-		let mut items_at_prio = self
-			.items
-			.remove(&prio)
-			.unwrap_or_else(|| VecDeque::with_capacity(4));
-		items_at_prio.push_back(item);
-		self.items.insert(prio, items_at_prio);
-	}
-	fn pop(&mut self) -> Option<SendQueueItem> {
-		match self.items.pop_first() {
-			None => None,
-			Some((prio, mut items_at_prio)) => {
-				let ret = items_at_prio.pop_front();
-				if !items_at_prio.is_empty() {
-					self.items.insert(prio, items_at_prio);
-				}
-				ret.or_else(|| self.pop())
-			}
-		}
-	}
-	fn is_empty(&self) -> bool {
-		self.items.iter().all(|(_k, v)| v.is_empty())
-	}
-}
-
-#[async_trait]
-pub(crate) trait SendLoop: Sync {
-	async fn send_loop<W>(
-		self: Arc<Self>,
-		mut msg_recv: mpsc::UnboundedReceiver<Option<(RequestID, RequestPriority, Vec<u8>)>>,
-		mut write: W,
-	) -> Result<(), Error>
-	where
-		W: AsyncWriteExt + Unpin + Send + Sync,
-	{
-		let mut sending = SendQueue::new();
-		let mut should_exit = false;
-		while !should_exit || !sending.is_empty() {
-			if let Ok(sth) = msg_recv.try_recv() {
-				if let Some((id, prio, data)) = sth {
-					trace!("send_loop: got {}, {} bytes", id, data.len());
-					sending.push(SendQueueItem {
-						id,
-						prio,
-						data,
-						cursor: 0,
-					});
-				} else {
-					should_exit = true;
-				}
-			} else if let Some(mut item) = sending.pop() {
-				trace!(
-					"send_loop: sending bytes for {} ({} bytes, {} already sent)",
-					item.id,
-					item.data.len(),
-					item.cursor
-				);
-				let header_id = u16::to_be_bytes(item.id);
-				write.write_all(&header_id[..]).await?;
-
-				if item.data.len() - item.cursor > MAX_CHUNK_SIZE {
-					let header_size = u16::to_be_bytes(MAX_CHUNK_SIZE as u16 | 0x8000);
-					write.write_all(&header_size[..]).await?;
-
-					let new_cursor = item.cursor + MAX_CHUNK_SIZE as usize;
-					write.write_all(&item.data[item.cursor..new_cursor]).await?;
-					item.cursor = new_cursor;
-
-					sending.push(item);
-				} else {
-					let send_len = (item.data.len() - item.cursor) as u16;
-
-					let header_size = u16::to_be_bytes(send_len);
-					write.write_all(&header_size[..]).await?;
-
-					write.write_all(&item.data[item.cursor..]).await?;
-				}
-				write.flush().await?;
-			} else {
-				let sth = msg_recv
-					.recv()
-					.await
-					.ok_or_else(|| Error::Message("Connection closed.".into()))?;
-				if let Some((id, prio, data)) = sth {
-					trace!("send_loop: got {}, {} bytes", id, data.len());
-					sending.push(SendQueueItem {
-						id,
-						prio,
-						data,
-						cursor: 0,
-					});
-				} else {
-					should_exit = true;
-				}
-			}
-		}
-		Ok(())
-	}
-}
-
-#[async_trait]
-pub(crate) trait RecvLoop: Sync + 'static {
-	// Returns true if we should stop receiving after this
-	async fn recv_handler(self: Arc<Self>, id: RequestID, msg: Vec<u8>);
-
-	async fn recv_loop<R>(self: Arc<Self>, mut read: R) -> Result<(), Error>
-	where
-		R: AsyncReadExt + Unpin + Send + Sync,
-	{
-		let mut receiving = HashMap::new();
-		loop {
-			trace!("recv_loop: reading packet");
-			let mut header_id = [0u8; 2];
-			read.read_exact(&mut header_id[..]).await?;
-			let id = RequestID::from_be_bytes(header_id);
-			trace!("recv_loop: got header id: {:04x}", id);
-
-			let mut header_size = [0u8; 2];
-			read.read_exact(&mut header_size[..]).await?;
-			let size = RequestID::from_be_bytes(header_size);
-			trace!("recv_loop: got header size: {:04x}", size);
-
-			let has_cont = (size & 0x8000) != 0;
-			let size = size & !0x8000;
-
-			let mut next_slice = vec![0; size as usize];
-			read.read_exact(&mut next_slice[..]).await?;
-			trace!("recv_loop: read {} bytes", next_slice.len());
-
-			let mut msg_bytes: Vec<_> = receiving.remove(&id).unwrap_or_default();
-			msg_bytes.extend_from_slice(&next_slice[..]);
-
-			if has_cont {
-				receiving.insert(id, msg_bytes);
-			} else {
-				tokio::spawn(self.clone().recv_handler(id, msg_bytes));
-			}
-		}
-	}
-}

src/recv.rs

@@ -0,0 +1,153 @@
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use bytes::Bytes;
+use log::*;
+
+use futures::AsyncReadExt;
+use tokio::sync::mpsc;
+
+use crate::error::*;
+use crate::send::*;
+use crate::stream::*;
+
+/// Structure to warn when the sender is dropped before end of stream was reached, like when
+/// connection to some remote drops while transmitting data
+struct Sender {
+	inner: Option<mpsc::UnboundedSender<Packet>>,
+}
+
+impl Sender {
+	fn new(inner: mpsc::UnboundedSender<Packet>) -> Self {
+		Sender { inner: Some(inner) }
+	}
+
+	fn send(&self, packet: Packet) {
+		let _ = self.inner.as_ref().unwrap().send(packet);
+	}
+
+	fn end(&mut self) {
+		self.inner = None;
+	}
+}
+
+impl Drop for Sender {
+	fn drop(&mut self) {
+		if let Some(inner) = self.inner.take() {
+			let _ = inner.send(Err(std::io::Error::new(
+				std::io::ErrorKind::BrokenPipe,
+				"Netapp connection dropped before end of stream",
+			)));
+		}
+	}
+}
+
+/// The RecvLoop trait, which is implemented both by the client and the server
+/// connection objects (ServerConn and ClientConn) adds a method `.recv_loop()`
+/// and a prototype of a handler for received messages `.recv_handler()` that
+/// must be filled by implementors. `.recv_loop()` receives messages in a loop
+/// according to the protocol defined above: chunks of message in progress of being
+/// received are stored in a buffer, and when the last chunk of a message is received,
+/// the full message is passed to the receive handler.
+#[async_trait]
+pub(crate) trait RecvLoop: Sync + 'static {
+	fn recv_handler(self: &Arc<Self>, id: RequestID, stream: ByteStream);
+	fn cancel_handler(self: &Arc<Self>, _id: RequestID) {}
+
+	async fn recv_loop<R>(self: Arc<Self>, mut read: R, debug_name: String) -> Result<(), Error>
+	where
+		R: AsyncReadExt + Unpin + Send + Sync,
+	{
+		let mut streams: HashMap<RequestID, Sender> = HashMap::new();
+		loop {
+			trace!(
+				"recv_loop({}): in_progress = {:?}",
+				debug_name,
+				streams.iter().map(|(id, _)| id).collect::<Vec<_>>()
+			);
+
+			let mut header_id = [0u8; RequestID::BITS as usize / 8];
+			match read.read_exact(&mut header_id[..]).await {
+				Ok(_) => (),
+				Err(e) if e.kind() == std::io::ErrorKind::UnexpectedEof => break,
+				Err(e) => return Err(e.into()),
+			};
+			let id = RequestID::from_be_bytes(header_id);
+
+			let mut header_size = [0u8; ChunkLength::BITS as usize / 8];
+			read.read_exact(&mut header_size[..]).await?;
+			let size = ChunkLength::from_be_bytes(header_size);
+
+			if size == CANCEL_REQUEST {
+				if let Some(mut stream) = streams.remove(&id) {
+					let _ = stream.send(Err(std::io::Error::new(
+						std::io::ErrorKind::Other,
+						"netapp: cancel requested",
+					)));
+					stream.end();
+				}
+				self.cancel_handler(id);
+				continue;
+			}
+
+			let has_cont = (size & CHUNK_FLAG_HAS_CONTINUATION) != 0;
+			let is_error = (size & CHUNK_FLAG_ERROR) != 0;
+			let size = (size & CHUNK_LENGTH_MASK) as usize;
+			let mut next_slice = vec![0; size as usize];
+			read.read_exact(&mut next_slice[..]).await?;
+
+			let packet = if is_error {
+				let kind = u8_to_io_errorkind(next_slice[0]);
+				let msg =
+					std::str::from_utf8(&next_slice[1..]).unwrap_or("<invalid utf8 error message>");
+				debug!(
+					"recv_loop({}): got id {}, error {:?}: {}",
+					debug_name, id, kind, msg
+				);
+				Some(Err(std::io::Error::new(kind, msg.to_string())))
+			} else {
+				trace!(
+					"recv_loop({}): got id {}, size {}, has_cont {}",
+					debug_name,
+					id,
+					size,
+					has_cont
+				);
+				if !next_slice.is_empty() {
+					Some(Ok(Bytes::from(next_slice)))
+				} else {
+					None
+				}
+			};
+
+			let mut sender = if let Some(send) = streams.remove(&(id)) {
+				send
+			} else {
+				let (send, recv) = mpsc::unbounded_channel();
+				trace!("recv_loop({}): id {} is new channel", debug_name, id);
+				self.recv_handler(
+					id,
+					Box::pin(tokio_stream::wrappers::UnboundedReceiverStream::new(recv)),
+				);
+				Sender::new(send)
+			};
+
+			if let Some(packet) = packet {
+				// If we cannot put packet in channel, it means that the
+				// receiving end of the channel is disconnected.
+				// We still need to reach eos before dropping this sender
+				let _ = sender.send(packet);
+			}
+
+			if has_cont {
+				assert!(!is_error);
+				streams.insert(id, sender);
+			} else {
+				trace!("recv_loop({}): close channel id {}", debug_name, id);
+				sender.end();
+			}
+		}
+		Ok(())
+	}
+}

src/send.rs

@@ -0,0 +1,356 @@
+use std::collections::{HashMap, VecDeque};
+use std::pin::Pin;
+use std::sync::Arc;
+use std::task::{Context, Poll};
+
+use async_trait::async_trait;
+use bytes::{BufMut, Bytes, BytesMut};
+use log::*;
+
+use futures::{AsyncWriteExt, Future};
+use kuska_handshake::async_std::BoxStreamWrite;
+use tokio::sync::mpsc;
+
+use crate::error::*;
+use crate::message::*;
+use crate::stream::*;
+
+// Messages are sent by chunks
+// Chunk format:
+// - u32 BE: request id (same for request and response)
+// - u16 BE: chunk length + flags:
+//		CHUNK_FLAG_HAS_CONTINUATION when this is not the last chunk of the stream
+//		CHUNK_FLAG_ERROR if this chunk denotes an error
+//		(these two flags are exclusive, an error denotes the end of the stream)
+//		**special value** 0xFFFF indicates a CANCEL message
+// - [u8; chunk_length], either
+//   - if not error: chunk data
+//   - if error:
+//       - u8: error kind, encoded using error::io_errorkind_to_u8
+//       - rest: error message
+//   - absent for cancel messag
+
+pub(crate) type RequestID = u32;
+pub(crate) type ChunkLength = u16;
+
+pub(crate) const MAX_CHUNK_LENGTH: ChunkLength = 0x3FF0;
+pub(crate) const CHUNK_FLAG_ERROR: ChunkLength = 0x4000;
+pub(crate) const CHUNK_FLAG_HAS_CONTINUATION: ChunkLength = 0x8000;
+pub(crate) const CHUNK_LENGTH_MASK: ChunkLength = 0x3FFF;
+pub(crate) const CANCEL_REQUEST: ChunkLength = 0xFFFF;
+
+pub(crate) enum SendItem {
+	Stream(RequestID, RequestPriority, Option<OrderTag>, ByteStream),
+	Cancel(RequestID),
+}
+
+// ----
+
+struct SendQueue {
+	items: Vec<(u8, SendQueuePriority)>,
+}
+
+struct SendQueuePriority {
+	items: VecDeque<SendQueueItem>,
+	order: HashMap<u64, VecDeque<u64>>,
+}
+
+struct SendQueueItem {
+	id: RequestID,
+	prio: RequestPriority,
+	order_tag: Option<OrderTag>,
+	data: ByteStreamReader,
+	sent: usize,
+}
+
+impl SendQueue {
+	fn new() -> Self {
+		Self {
+			items: Vec::with_capacity(64),
+		}
+	}
+	fn push(&mut self, item: SendQueueItem) {
+		let prio = item.prio;
+		let pos_prio = match self.items.binary_search_by(|(p, _)| p.cmp(&prio)) {
+			Ok(i) => i,
+			Err(i) => {
+				self.items.insert(i, (prio, SendQueuePriority::new()));
+				i
+			}
+		};
+		self.items[pos_prio].1.push(item);
+	}
+	fn remove(&mut self, id: RequestID) {
+		for (_, prioq) in self.items.iter_mut() {
+			prioq.remove(id);
+		}
+		self.items.retain(|(_prio, q)| !q.is_empty());
+	}
+	fn is_empty(&self) -> bool {
+		self.items.iter().all(|(_k, v)| v.is_empty())
+	}
+
+	// this is like an async fn, but hand implemented
+	fn next_ready(&mut self) -> SendQueuePollNextReady<'_> {
+		SendQueuePollNextReady { queue: self }
+	}
+}
+
+impl SendQueuePriority {
+	fn new() -> Self {
+		Self {
+			items: VecDeque::new(),
+			order: HashMap::new(),
+		}
+	}
+	fn push(&mut self, item: SendQueueItem) {
+		if let Some(OrderTag(stream, order)) = item.order_tag {
+			let order_vec = self.order.entry(stream).or_default();
+			let i = order_vec.iter().take_while(|o2| **o2 < order).count();
+			order_vec.insert(i, order);
+		}
+		self.items.push_front(item);
+	}
+	fn remove(&mut self, id: RequestID) {
+		if let Some(i) = self.items.iter().position(|x| x.id == id) {
+			let item = self.items.remove(i).unwrap();
+			if let Some(OrderTag(stream, order)) = item.order_tag {
+				let order_vec = self.order.get_mut(&stream).unwrap();
+				let j = order_vec.iter().position(|x| *x == order).unwrap();
+				order_vec.remove(j).unwrap();
+				if order_vec.is_empty() {
+					self.order.remove(&stream);
+				}
+			}
+		}
+	}
+	fn is_empty(&self) -> bool {
+		self.items.is_empty()
+	}
+	fn poll_next_ready(&mut self, ctx: &mut Context<'_>) -> Poll<(RequestID, DataFrame)> {
+		for (j, item) in self.items.iter_mut().enumerate() {
+			if let Some(OrderTag(stream, order)) = item.order_tag {
+				if order > *self.order.get(&stream).unwrap().front().unwrap() {
+					continue;
+				}
+			}
+
+			let mut item_reader = item.data.read_exact_or_eos(MAX_CHUNK_LENGTH as usize);
+			if let Poll::Ready(bytes_or_err) = Pin::new(&mut item_reader).poll(ctx) {
+				let id = item.id;
+				let eos = item.data.eos();
+
+				let packet = bytes_or_err.map_err(|e| match e {
+					ReadExactError::Stream(err) => err,
+					_ => unreachable!(),
+				});
+
+				let is_err = packet.is_err();
+				let data_frame = DataFrame::from_packet(packet, !eos);
+				item.sent += data_frame.data().len();
+
+				if eos || is_err {
+					// If item had an order tag, remove it from the corresponding ordering list
+					if let Some(OrderTag(stream, order)) = item.order_tag {
+						let order_stream = self.order.get_mut(&stream).unwrap();
+						assert_eq!(order_stream.pop_front(), Some(order));
+						if order_stream.is_empty() {
+							self.order.remove(&stream);
+						}
+					}
+					// Remove item from sending queue
+					self.items.remove(j);
+				} else {
+					// Move item later in send queue to implement LAS scheduling
+					// (LAS = Least Attained Service)
+					for k in j..self.items.len() - 1 {
+						if self.items[k].sent >= self.items[k + 1].sent {
+							self.items.swap(k, k + 1);
+						} else {
+							break;
+						}
+					}
+				}
+
+				return Poll::Ready((id, data_frame));
+			}
+		}
+
+		Poll::Pending
+	}
+	fn dump(&self, prio: u8) -> String {
+		self.items
+			.iter()
+			.map(|i| format!("[{} {} {:?} @{}]", prio, i.id, i.order_tag, i.sent))
+			.collect::<Vec<_>>()
+			.join(" ")
+	}
+}
+
+struct SendQueuePollNextReady<'a> {
+	queue: &'a mut SendQueue,
+}
+
+impl<'a> futures::Future for SendQueuePollNextReady<'a> {
+	type Output = (RequestID, DataFrame);
+
+	fn poll(mut self: Pin<&mut Self>, ctx: &mut Context<'_>) -> Poll<Self::Output> {
+		for (i, (_prio, items_at_prio)) in self.queue.items.iter_mut().enumerate() {
+			if let Poll::Ready(res) = items_at_prio.poll_next_ready(ctx) {
+				if items_at_prio.is_empty() {
+					self.queue.items.remove(i);
+				}
+				return Poll::Ready(res);
+			}
+		}
+		// If the queue is empty, this futures is eternally pending.
+		// This is ok because we use it in a select with another future
+		// that can interrupt it.
+		Poll::Pending
+	}
+}
+
+enum DataFrame {
+	/// a fixed size buffer containing some data + a boolean indicating whether
+	/// there may be more data comming from this stream. Can be used for some
+	/// optimization. It's an error to set it to false if there is more data, but it is correct
+	/// (albeit sub-optimal) to set it to true if there is nothing coming after
+	Data(Bytes, bool),
+	/// An error code automatically signals the end of the stream
+	Error(Bytes),
+}
+
+impl DataFrame {
+	fn from_packet(p: Packet, has_cont: bool) -> Self {
+		match p {
+			Ok(bytes) => {
+				assert!(bytes.len() <= MAX_CHUNK_LENGTH as usize);
+				Self::Data(bytes, has_cont)
+			}
+			Err(e) => {
+				let mut buf = BytesMut::new();
+				buf.put_u8(io_errorkind_to_u8(e.kind()));
+
+				let msg = format!("{}", e).into_bytes();
+				if msg.len() > (MAX_CHUNK_LENGTH - 1) as usize {
+					buf.put(&msg[..(MAX_CHUNK_LENGTH - 1) as usize]);
+				} else {
+					buf.put(&msg[..]);
+				}
+
+				Self::Error(buf.freeze())
+			}
+		}
+	}
+
+	fn header(&self) -> [u8; 2] {
+		let header_u16 = match self {
+			DataFrame::Data(data, false) => data.len() as u16,
+			DataFrame::Data(data, true) => data.len() as u16 | CHUNK_FLAG_HAS_CONTINUATION,
+			DataFrame::Error(msg) => msg.len() as u16 | CHUNK_FLAG_ERROR,
+		};
+		ChunkLength::to_be_bytes(header_u16)
+	}
+
+	fn data(&self) -> &[u8] {
+		match self {
+			DataFrame::Data(ref data, _) => &data[..],
+			DataFrame::Error(ref msg) => &msg[..],
+		}
+	}
+}
+
+/// The SendLoop trait, which is implemented both by the client and the server
+/// connection objects (ServerConna and ClientConn) adds a method `.send_loop()`
+/// that takes a channel of messages to send and an asynchronous writer,
+/// and sends messages from the channel to the async writer, putting them in a queue
+/// before being sent and doing the round-robin sending strategy.
+///
+/// The `.send_loop()` exits when the sending end of the channel is closed,
+/// or if there is an error at any time writing to the async writer.
+#[async_trait]
+pub(crate) trait SendLoop: Sync {
+	async fn send_loop<W>(
+		self: Arc<Self>,
+		msg_recv: mpsc::UnboundedReceiver<SendItem>,
+		mut write: BoxStreamWrite<W>,
+		debug_name: String,
+	) -> Result<(), Error>
+	where
+		W: AsyncWriteExt + Unpin + Send + Sync,
+	{
+		let mut sending = SendQueue::new();
+		let mut msg_recv = Some(msg_recv);
+		while msg_recv.is_some() || !sending.is_empty() {
+			trace!(
+				"send_loop({}): queue = {:?}",
+				debug_name,
+				sending
+					.items
+					.iter()
+					.map(|(prio, i)| i.dump(*prio))
+					.collect::<Vec<_>>()
+					.join(" ; ")
+			);
+
+			let recv_fut = async {
+				if let Some(chan) = &mut msg_recv {
+					chan.recv().await
+				} else {
+					futures::future::pending().await
+				}
+			};
+			let send_fut = sending.next_ready();
+
+			// recv_fut is cancellation-safe according to tokio doc,
+			// send_fut is cancellation-safe as implemented above?
+			tokio::select! {
+				biased;	// always read incomming channel first if it has data
+				sth = recv_fut => {
+					match sth {
+						Some(SendItem::Stream(id, prio, order_tag, data)) => {
+							trace!("send_loop({}): add stream {} to send", debug_name, id);
+							sending.push(SendQueueItem {
+								id,
+								prio,
+								order_tag,
+								data: ByteStreamReader::new(data),
+								sent: 0,
+							})
+						}
+						Some(SendItem::Cancel(id)) => {
+							trace!("send_loop({}): cancelling {}", debug_name, id);
+							sending.remove(id);
+							let header_id = RequestID::to_be_bytes(id);
+							write.write_all(&header_id[..]).await?;
+							write.write_all(&ChunkLength::to_be_bytes(CANCEL_REQUEST)).await?;
+							write.flush().await?;
+						}
+						None => {
+							msg_recv = None;
+						}
+					};
+				}
+				(id, data) = send_fut => {
+					trace!(
+						"send_loop({}): id {}, send {} bytes, header_size {}",
+						debug_name,
+						id,
+						data.data().len(),
+						hex::encode(data.header())
+					);
+
+					let header_id = RequestID::to_be_bytes(id);
+					write.write_all(&header_id[..]).await?;
+
+					write.write_all(&data.header()).await?;
+					write.write_all(data.data()).await?;
+					write.flush().await?;
+				}
+			}
+		}
+
+		let _ = write.goodbye().await;
+		Ok(())
+	}
+}

src/server.rs

@@ -0,0 +1,222 @@
+use std::collections::HashMap;
+use std::net::SocketAddr;
+use std::sync::{Arc, Mutex};
+
+use arc_swap::ArcSwapOption;
+use async_trait::async_trait;
+use log::*;
+
+use futures::io::{AsyncReadExt, AsyncWriteExt};
+use kuska_handshake::async_std::{handshake_server, BoxStream};
+use tokio::net::TcpStream;
+use tokio::select;
+use tokio::sync::{mpsc, watch};
+use tokio_util::compat::*;
+
+#[cfg(feature = "telemetry")]
+use opentelemetry::{
+	trace::{FutureExt, Span, SpanKind, TraceContextExt, TraceId, Tracer},
+	Context, KeyValue,
+};
+#[cfg(feature = "telemetry")]
+use opentelemetry_contrib::trace::propagator::binary::*;
+#[cfg(feature = "telemetry")]
+use rand::{thread_rng, Rng};
+
+use crate::error::*;
+use crate::message::*;
+use crate::netapp::*;
+use crate::recv::*;
+use crate::send::*;
+use crate::stream::*;
+use crate::util::*;
+
+// The client and server connection structs (client.rs and server.rs)
+// build upon the chunking mechanism which is exclusively contained
+// in proto.rs.
+// Here, we just care about sending big messages without size limit.
+// The format of these messages is described below.
+// Chunking happens independently.
+
+// Request message format (client -> server):
+// - u8 priority
+// - u8 path length
+// - [u8; path length] path
+// - [u8; *] data
+
+// Response message format (server -> client):
+// - u8 response code
+// - [u8; *] response
+
+pub(crate) struct ServerConn {
+	pub(crate) remote_addr: SocketAddr,
+	pub(crate) peer_id: NodeID,
+
+	netapp: Arc<NetApp>,
+
+	resp_send: ArcSwapOption<mpsc::UnboundedSender<SendItem>>,
+	running_handlers: Mutex<HashMap<RequestID, tokio::task::JoinHandle<()>>>,
+}
+
+impl ServerConn {
+	pub(crate) async fn run(
+		netapp: Arc<NetApp>,
+		socket: TcpStream,
+		must_exit: watch::Receiver<bool>,
+	) -> Result<(), Error> {
+		let remote_addr = socket.peer_addr()?;
+		let mut socket = socket.compat();
+
+		// Do handshake to authenticate client
+		let handshake = handshake_server(
+			&mut socket,
+			netapp.netid.clone(),
+			netapp.id,
+			netapp.privkey.clone(),
+		)
+		.await?;
+		let peer_id = handshake.peer_pk;
+
+		debug!(
+			"Handshake complete (server) with {}@{}",
+			hex::encode(peer_id),
+			remote_addr
+		);
+
+		// Create BoxStream layer that encodes content
+		let (read, write) = socket.split();
+		let (read, mut write) =
+			BoxStream::from_handshake(read, write, handshake, 0x8000).split_read_write();
+
+		// Before doing anything, send version tag, so that client
+		// can check and disconnect if version is wrong
+		write.write_all(&netapp.version_tag[..]).await?;
+		write.flush().await?;
+
+		// Build and launch stuff that handles requests server-side
+		let (resp_send, resp_recv) = mpsc::unbounded_channel();
+
+		let conn = Arc::new(ServerConn {
+			netapp: netapp.clone(),
+			remote_addr,
+			peer_id,
+			resp_send: ArcSwapOption::new(Some(Arc::new(resp_send))),
+			running_handlers: Mutex::new(HashMap::new()),
+		});
+
+		netapp.connected_as_server(peer_id, conn.clone());
+
+		let debug_name = format!("SRV {}", hex::encode(&peer_id[..8]));
+		let debug_name_2 = debug_name.clone();
+
+		let conn2 = conn.clone();
+		let recv_future = tokio::spawn(async move {
+			select! {
+				r = conn2.recv_loop(read, debug_name_2) => r,
+				_ = await_exit(must_exit) => Ok(())
+			}
+		});
+		let send_future = tokio::spawn(conn.clone().send_loop(resp_recv, write, debug_name));
+
+		recv_future.await.log_err("ServerConn recv_loop");
+		conn.resp_send.store(None);
+		send_future.await.log_err("ServerConn send_loop");
+
+		netapp.disconnected_as_server(&peer_id, conn);
+
+		Ok(())
+	}
+
+	async fn recv_handler_aux(self: &Arc<Self>, req_enc: ReqEnc) -> Result<RespEnc, Error> {
+		let path = String::from_utf8(req_enc.path.to_vec())?;
+
+		let handler_opt = {
+			let endpoints = self.netapp.endpoints.read().unwrap();
+			endpoints.get(&path).map(|e| e.clone_endpoint())
+		};
+
+		if let Some(handler) = handler_opt {
+			cfg_if::cfg_if! {
+				if #[cfg(feature = "telemetry")] {
+					let tracer = opentelemetry::global::tracer("netapp");
+
+					let mut span = if !req_enc.telemetry_id.is_empty() {
+						let propagator = BinaryPropagator::new();
+						let context = propagator.from_bytes(req_enc.telemetry_id.to_vec());
+						let context = Context::new().with_remote_span_context(context);
+						tracer.span_builder(format!(">> RPC {}", path))
+							.with_kind(SpanKind::Server)
+							.start_with_context(&tracer, &context)
+					} else {
+						let mut rng = thread_rng();
+						let trace_id = TraceId::from_bytes(rng.gen());
+						tracer
+							.span_builder(format!(">> RPC {}", path))
+							.with_kind(SpanKind::Server)
+							.with_trace_id(trace_id)
+							.start(&tracer)
+					};
+					span.set_attribute(KeyValue::new("path", path.to_string()));
+					span.set_attribute(KeyValue::new("len_query_msg", req_enc.msg.len() as i64));
+
+					handler.handle(req_enc, self.peer_id)
+						.with_context(Context::current_with_span(span))
+						.await
+				} else {
+					handler.handle(req_enc, self.peer_id).await
+				}
+			}
+		} else {
+			Err(Error::NoHandler)
+		}
+	}
+}
+
+impl SendLoop for ServerConn {}
+
+#[async_trait]
+impl RecvLoop for ServerConn {
+	fn recv_handler(self: &Arc<Self>, id: RequestID, stream: ByteStream) {
+		let resp_send = match self.resp_send.load_full() {
+			Some(c) => c,
+			None => return,
+		};
+
+		let mut rh = self.running_handlers.lock().unwrap();
+
+		let self2 = self.clone();
+		let jh = tokio::spawn(async move {
+			debug!("server: recv_handler got {}", id);
+
+			let (prio, resp_enc_result) = match ReqEnc::decode(stream).await {
+				Ok(req_enc) => (req_enc.prio, self2.recv_handler_aux(req_enc).await),
+				Err(e) => (PRIO_HIGH, Err(e)),
+			};
+
+			debug!("server: sending response to {}", id);
+
+			let (resp_stream, resp_order) = RespEnc::encode(resp_enc_result);
+			resp_send
+				.send(SendItem::Stream(id, prio, resp_order, resp_stream))
+				.log_err("ServerConn recv_handler send resp bytes");
+
+			self2.running_handlers.lock().unwrap().remove(&id);
+		});
+
+		rh.insert(id, jh);
+	}
+
+	fn cancel_handler(self: &Arc<Self>, id: RequestID) {
+		trace!("received cancel for request {}", id);
+
+		// If the handler is still running, abort it now
+		if let Some(jh) = self.running_handlers.lock().unwrap().remove(&id) {
+			jh.abort();
+		}
+
+		// Inform the response sender that we don't need to send the response
+		if let Some(resp_send) = self.resp_send.load_full() {
+			let _ = resp_send.send(SendItem::Cancel(id));
+		}
+	}
+}

src/stream.rs

@@ -0,0 +1,202 @@
+use std::pin::Pin;
+use std::task::{Context, Poll};
+
+use bytes::Bytes;
+
+use futures::Future;
+use futures::{Stream, StreamExt};
+use tokio::io::AsyncRead;
+
+use crate::bytes_buf::BytesBuf;
+
+/// A stream of bytes (click to read more).
+///
+/// When sent through Netapp, the Vec may be split in smaller chunk in such a way
+/// consecutive Vec may get merged, but Vec and error code may not be reordered
+///
+/// Items sent in the ByteStream may be errors of type `std::io::Error`.
+/// An error indicates the end of the ByteStream: a reader should no longer read
+/// after recieving an error, and a writer should stop writing after sending an error.
+pub type ByteStream = Pin<Box<dyn Stream<Item = Packet> + Send + Sync>>;
+
+/// A packet sent in a ByteStream, which may contain either
+/// a Bytes object or an error
+pub type Packet = Result<Bytes, std::io::Error>;
+
+// ----
+
+/// A helper struct to read defined lengths of data from a BytesStream
+pub struct ByteStreamReader {
+	stream: ByteStream,
+	buf: BytesBuf,
+	eos: bool,
+	err: Option<std::io::Error>,
+}
+
+impl ByteStreamReader {
+	/// Creates a new `ByteStreamReader` from a `ByteStream`
+	pub fn new(stream: ByteStream) -> Self {
+		ByteStreamReader {
+			stream,
+			buf: BytesBuf::new(),
+			eos: false,
+			err: None,
+		}
+	}
+
+	/// Read exactly `read_len` bytes from the underlying stream
+	/// (returns a future)
+	pub fn read_exact(&mut self, read_len: usize) -> ByteStreamReadExact<'_> {
+		ByteStreamReadExact {
+			reader: self,
+			read_len,
+			fail_on_eos: true,
+		}
+	}
+
+	/// Read at most `read_len` bytes from the underlying stream, or less
+	/// if the end of the stream is reached (returns a future)
+	pub fn read_exact_or_eos(&mut self, read_len: usize) -> ByteStreamReadExact<'_> {
+		ByteStreamReadExact {
+			reader: self,
+			read_len,
+			fail_on_eos: false,
+		}
+	}
+
+	/// Read exactly one byte from the underlying stream and returns it
+	/// as an u8
+	pub async fn read_u8(&mut self) -> Result<u8, ReadExactError> {
+		Ok(self.read_exact(1).await?[0])
+	}
+
+	/// Read exactly two bytes from the underlying stream and returns them as an u16 (using
+	/// big-endian decoding)
+	pub async fn read_u16(&mut self) -> Result<u16, ReadExactError> {
+		let bytes = self.read_exact(2).await?;
+		let mut b = [0u8; 2];
+		b.copy_from_slice(&bytes[..]);
+		Ok(u16::from_be_bytes(b))
+	}
+
+	/// Read exactly four bytes from the underlying stream and returns them as an u32 (using
+	/// big-endian decoding)
+	pub async fn read_u32(&mut self) -> Result<u32, ReadExactError> {
+		let bytes = self.read_exact(4).await?;
+		let mut b = [0u8; 4];
+		b.copy_from_slice(&bytes[..]);
+		Ok(u32::from_be_bytes(b))
+	}
+
+	/// Transforms the stream reader back into the underlying stream (starting
+	/// after everything that the reader has read)
+	pub fn into_stream(self) -> ByteStream {
+		let buf_stream = futures::stream::iter(self.buf.into_slices().into_iter().map(Ok));
+		if let Some(err) = self.err {
+			Box::pin(buf_stream.chain(futures::stream::once(async move { Err(err) })))
+		} else if self.eos {
+			Box::pin(buf_stream)
+		} else {
+			Box::pin(buf_stream.chain(self.stream))
+		}
+	}
+
+	/// Tries to fill the internal read buffer from the underlying stream if it is empty.
+	/// Calling this might be necessary to ensure that `.eos()` returns a correct
+	/// result, otherwise the reader might not be aware that the underlying
+	/// stream has nothing left to return.
+	pub async fn fill_buffer(&mut self) {
+		if self.buf.is_empty() {
+			let packet = self.stream.next().await;
+			self.add_stream_next(packet);
+		}
+	}
+
+	/// Clears the internal read buffer and returns its content
+	pub fn take_buffer(&mut self) -> Bytes {
+		self.buf.take_all()
+	}
+
+	/// Returns true if the end of the underlying stream has been reached
+	pub fn eos(&self) -> bool {
+		self.buf.is_empty() && self.eos
+	}
+
+	fn try_get(&mut self, read_len: usize) -> Option<Bytes> {
+		self.buf.take_exact(read_len)
+	}
+
+	fn add_stream_next(&mut self, packet: Option<Packet>) {
+		match packet {
+			Some(Ok(slice)) => {
+				self.buf.extend(slice);
+			}
+			Some(Err(e)) => {
+				self.err = Some(e);
+				self.eos = true;
+			}
+			None => {
+				self.eos = true;
+			}
+		}
+	}
+}
+
+/// The error kind that can be returned by `ByteStreamReader::read_exact` and
+/// `ByteStreamReader::read_exact_or_eos`
+pub enum ReadExactError {
+	/// The end of the stream was reached before the requested number of bytes could be read
+	UnexpectedEos,
+	/// The underlying data stream returned an IO error when trying to read
+	Stream(std::io::Error),
+}
+
+/// The future returned by `ByteStreamReader::read_exact` and
+/// `ByteStreamReader::read_exact_or_eos`
+#[pin_project::pin_project]
+pub struct ByteStreamReadExact<'a> {
+	#[pin]
+	reader: &'a mut ByteStreamReader,
+	read_len: usize,
+	fail_on_eos: bool,
+}
+
+impl<'a> Future for ByteStreamReadExact<'a> {
+	type Output = Result<Bytes, ReadExactError>;
+
+	fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<Bytes, ReadExactError>> {
+		let mut this = self.project();
+
+		loop {
+			if let Some(bytes) = this.reader.try_get(*this.read_len) {
+				return Poll::Ready(Ok(bytes));
+			}
+			if let Some(err) = &this.reader.err {
+				let err = std::io::Error::new(err.kind(), format!("{}", err));
+				return Poll::Ready(Err(ReadExactError::Stream(err)));
+			}
+			if this.reader.eos {
+				if *this.fail_on_eos {
+					return Poll::Ready(Err(ReadExactError::UnexpectedEos));
+				} else {
+					return Poll::Ready(Ok(this.reader.take_buffer()));
+				}
+			}
+
+			let next_packet = futures::ready!(this.reader.stream.as_mut().poll_next(cx));
+			this.reader.add_stream_next(next_packet);
+		}
+	}
+}
+
+// ----
+
+/// Turns a `tokio::io::AsyncRead` asynchronous reader into a `ByteStream`
+pub fn asyncread_stream<R: AsyncRead + Send + Sync + 'static>(reader: R) -> ByteStream {
+	Box::pin(tokio_util::io::ReaderStream::new(reader))
+}
+
+/// Turns a `ByteStream` into a `tokio::io::AsyncRead` asynchronous reader
+pub fn stream_asyncread(stream: ByteStream) -> impl AsyncRead + Send + Sync + 'static {
+	tokio_util::io::StreamReader::new(stream)
+}

src/test.rs

@@ -0,0 +1,118 @@
+use std::net::SocketAddr;
+use std::sync::Arc;
+use std::time::Duration;
+
+use tokio::select;
+use tokio::sync::watch;
+
+use sodiumoxide::crypto::auth;
+use sodiumoxide::crypto::sign::ed25519;
+
+use crate::netapp::*;
+use crate::peering::fullmesh::*;
+use crate::NodeID;
+
+#[tokio::test(flavor = "current_thread")]
+async fn test_with_basic_scheduler() {
+	env_logger::init();
+	run_test().await
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn test_with_threaded_scheduler() {
+	run_test().await
+}
+
+async fn run_test() {
+	select! {
+		_ = run_test_inner() => (),
+		_ = tokio::time::sleep(Duration::from_secs(20)) => panic!("timeout"),
+	}
+}
+
+async fn run_test_inner() {
+	let netid = auth::gen_key();
+
+	let (pk1, sk1) = ed25519::gen_keypair();
+	let (pk2, sk2) = ed25519::gen_keypair();
+	let (pk3, sk3) = ed25519::gen_keypair();
+
+	let addr1: SocketAddr = "127.0.0.1:19991".parse().unwrap();
+	let addr2: SocketAddr = "127.0.0.1:19992".parse().unwrap();
+	let addr3: SocketAddr = "127.0.0.1:19993".parse().unwrap();
+
+	let (stop_tx, stop_rx) = watch::channel(false);
+
+	let (thread1, _netapp1, peering1) =
+		run_netapp(netid.clone(), pk1, sk1, addr1, vec![], stop_rx.clone());
+	tokio::time::sleep(Duration::from_secs(2)).await;
+
+	// Connect second node and check it peers with everyone
+	let (thread2, _netapp2, peering2) = run_netapp(
+		netid.clone(),
+		pk2,
+		sk2,
+		addr2,
+		vec![(pk1, addr1)],
+		stop_rx.clone(),
+	);
+	tokio::time::sleep(Duration::from_secs(5)).await;
+
+	let pl1 = peering1.get_peer_list();
+	println!("A pl1: {:?}", pl1);
+	assert_eq!(pl1.len(), 2);
+
+	let pl2 = peering2.get_peer_list();
+	println!("A pl2: {:?}", pl2);
+	assert_eq!(pl2.len(), 2);
+
+	// Connect third ndoe and check it peers with everyone
+	let (thread3, _netapp3, peering3) =
+		run_netapp(netid, pk3, sk3, addr3, vec![(pk2, addr2)], stop_rx.clone());
+	tokio::time::sleep(Duration::from_secs(5)).await;
+
+	let pl1 = peering1.get_peer_list();
+	println!("B pl1: {:?}", pl1);
+	assert_eq!(pl1.len(), 3);
+
+	let pl2 = peering2.get_peer_list();
+	println!("B pl2: {:?}", pl2);
+	assert_eq!(pl2.len(), 3);
+
+	let pl3 = peering3.get_peer_list();
+	println!("B pl3: {:?}", pl3);
+	assert_eq!(pl3.len(), 3);
+
+	// Send stop signal and wait for everyone to finish
+	stop_tx.send(true).unwrap();
+	thread1.await.unwrap();
+	thread2.await.unwrap();
+	thread3.await.unwrap();
+}
+
+fn run_netapp(
+	netid: auth::Key,
+	_pk: NodeID,
+	sk: ed25519::SecretKey,
+	listen_addr: SocketAddr,
+	bootstrap_peers: Vec<(NodeID, SocketAddr)>,
+	must_exit: watch::Receiver<bool>,
+) -> (
+	tokio::task::JoinHandle<()>,
+	Arc<NetApp>,
+	Arc<FullMeshPeeringStrategy>,
+) {
+	let netapp = NetApp::new(0u64, netid, sk);
+	let peering = FullMeshPeeringStrategy::new(netapp.clone(), bootstrap_peers, None);
+
+	let peering2 = peering.clone();
+	let netapp2 = netapp.clone();
+	let fut = tokio::spawn(async move {
+		tokio::join!(
+			netapp2.listen(listen_addr, None, must_exit.clone()),
+			peering2.run(must_exit.clone()),
+		);
+	});
+
+	(fut, netapp, peering)
+}

src/util.rs

@@ -1,8 +1,11 @@
+use std::net::SocketAddr;
+
+use log::info;
 use serde::Serialize;
 
 use tokio::sync::watch;
 
-pub type NodeID = sodiumoxide::crypto::sign::ed25519::PublicKey;
+use crate::netapp::*;
 
 /// Utility function: encodes any serializable value in MessagePack binary format
 /// using the RMP library.
@@ -23,9 +26,10 @@ where
 
 /// This async function returns only when a true signal was received
 /// from a watcher that tells us when to exit.
+///
 /// Usefull in a select statement to interrupt another
 /// future:
-/// ```
+/// ```ignore
 /// select!(
 ///     _ = a_long_task() => Success,
 ///     _ = await_exit(must_exit) => Interrupted,
@@ -38,3 +42,57 @@ pub async fn await_exit(mut must_exit: watch::Receiver<bool>) {
 		}
 	}
 }
+
+/// Creates a watch that contains `false`, and that changes
+/// to `true` when a Ctrl+C signal is received.
+pub fn watch_ctrl_c() -> watch::Receiver<bool> {
+	let (send_cancel, watch_cancel) = watch::channel(false);
+	tokio::spawn(async move {
+		tokio::signal::ctrl_c()
+			.await
+			.expect("failed to install CTRL+C signal handler");
+		info!("Received CTRL+C, shutting down.");
+		send_cancel.send(true).unwrap();
+	});
+	watch_cancel
+}
+
+/// Parse a peer's address including public key, written in the format:
+/// `<public key hex>@<ip>:<port>`
+pub fn parse_peer_addr(peer: &str) -> Option<(NodeID, SocketAddr)> {
+	let delim = peer.find('@')?;
+	let (key, ip) = peer.split_at(delim);
+	let pubkey = NodeID::from_slice(&hex::decode(key).ok()?)?;
+	let ip = ip[1..].parse::<SocketAddr>().ok()?;
+	Some((pubkey, ip))
+}
+
+/// Parse and resolve a peer's address including public key, written in the format:
+/// `<public key hex>@<ip or hostname>:<port>`
+pub fn parse_and_resolve_peer_addr(peer: &str) -> Option<(NodeID, Vec<SocketAddr>)> {
+	use std::net::ToSocketAddrs;
+
+	let delim = peer.find('@')?;
+	let (key, host) = peer.split_at(delim);
+	let pubkey = NodeID::from_slice(&hex::decode(key).ok()?)?;
+	let hosts = host[1..].to_socket_addrs().ok()?.collect::<Vec<_>>();
+	if hosts.is_empty() {
+		return None;
+	}
+	Some((pubkey, hosts))
+}
+
+/// async version of parse_and_resolve_peer_addr
+pub async fn parse_and_resolve_peer_addr_async(peer: &str) -> Option<(NodeID, Vec<SocketAddr>)> {
+	let delim = peer.find('@')?;
+	let (key, host) = peer.split_at(delim);
+	let pubkey = NodeID::from_slice(&hex::decode(key).ok()?)?;
+	let hosts = tokio::net::lookup_host(&host[1..])
+		.await
+		.ok()?
+		.collect::<Vec<_>>();
+	if hosts.is_empty() {
+		return None;
+	}
+	Some((pubkey, hosts))
+}