Merge pull request 'add streaming body to requests and responses' (#3) from stream-body into mainv0.5.0

Reviewed-on: https://git.deuxfleurs.fr/lx/netapp/pulls/3

17 files changed, 1713 insertions, 619 deletions

diff --git a/src/bytes_buf.rs b/src/bytes_buf.rs
new file mode 100644
index 0000000..931be82
--- /dev/null
+++ b/src/bytes_buf.rs
@@ -0,0 +1,175 @@
+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.len() == 0 {
+			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();
+		if front.len() > len {
+			self.buf.push_front(front.slice(len..));
+			self.buf_len -= len;
+			front.slice(..len)
+		} else if front.len() == len {
+			self.buf_len -= len;
+			front
+		} else {
+			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 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());
+	}
+}
diff --git a/src/client.rs b/src/client.rs
index 5c5a05b..d82c91e 100644
--- a/src/client.rs
+++ b/src/client.rs
@@ -1,12 +1,18 @@
-use std::borrow::Borrow;
 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};
@@ -20,27 +26,22 @@ use opentelemetry::{
 #[cfg(feature = "telemetry")]
 use opentelemetry_contrib::trace::propagator::binary::*;
 
-use futures::io::AsyncReadExt;
-
-use async_trait::async_trait;
-
-use kuska_handshake::async_std::{handshake_client, BoxStream};
-
-use crate::endpoint::*;
 use crate::error::*;
+use crate::message::*;
 use crate::netapp::*;
-use crate::proto::*;
-use crate::proto2::*;
+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<(RequestID, RequestPriority, Vec<u8>)>>,
+	query_send: ArcSwapOption<mpsc::UnboundedSender<SendItem>>,
 
 	next_query_number: AtomicU32,
-	inflight: Mutex<HashMap<RequestID, oneshot::Sender<Vec<u8>>>>,
+	inflight: Mutex<HashMap<RequestID, oneshot::Sender<ByteStream>>>,
 }
 
 impl ClientConn {
@@ -139,15 +140,14 @@ impl ClientConn {
 		self.query_send.store(None);
 	}
 
-	pub(crate) async fn call<T, B>(
+	pub(crate) async fn call<T>(
 		self: Arc<Self>,
-		rq: B,
+		req: Req<T>,
 		path: &str,
 		prio: RequestPriority,
-	) -> Result<<T as Message>::Response, Error>
+	) -> Result<Resp<T>, Error>
 	where
 		T: Message,
-		B: Borrow<T>,
 	{
 		let query_send = self.query_send.load_full().ok_or(Error::ConnectionClosed)?;
 
@@ -162,24 +162,16 @@ impl ClientConn {
 					.with_kind(SpanKind::Client)
 					.start(&tracer);
 				let propagator = BinaryPropagator::new();
-				let telemetry_id = Some(propagator.to_bytes(span.span_context()).to_vec());
+				let telemetry_id: Bytes = propagator.to_bytes(span.span_context()).to_vec().into();
 			} else {
-				let telemetry_id: Option<Vec<u8>> = None;
+				let telemetry_id: Bytes = Bytes::new();
 			}
 		};
 
 		// Encode request
-		let body = rmp_to_vec_all_named(rq.borrow())?;
-		drop(rq);
-
-		let request = QueryMessage {
-			prio,
-			path: path.as_bytes(),
-			telemetry_id,
-			body: &body[..],
-		};
-		let bytes = request.encode();
-		drop(body);
+		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();
@@ -188,46 +180,41 @@ impl ClientConn {
 			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.");
-			}
+			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",
+				))
+			})));
 		}
 
-		trace!("request: query_send {}, {} bytes", id, bytes.len());
+		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", bytes.len() as i64));
+		span.set_attribute(KeyValue::new("len_query_msg", req_msg_len as i64));
 
-		query_send.send((id, prio, bytes))?;
+		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 resp = resp_recv
+				let stream = resp_recv
 					.with_context(Context::current_with_span(span))
 					.await?;
 			} else {
-				let resp = resp_recv.await?;
+				let stream = resp_recv.await?;
 			}
 		}
 
-		if resp.is_empty() {
-			return Err(Error::Message(
-				"Response is 0 bytes, either a collision or a protocol error".into(),
-			));
-		}
-
-		trace!("request response {}: ", id);
+		let stream = Box::pin(canceller.for_stream(stream));
 
-		let code = resp[0];
-		if code == 0 {
-			Ok(rmp_serde::decode::from_read_ref::<
-				_,
-				<T as Message>::Response,
-			>(&resp[1..])?)
-		} else {
-			let msg = String::from_utf8(resp[1..].to_vec()).unwrap_or_default();
-			Err(Error::Remote(code, msg))
-		}
+		let resp_enc = RespEnc::decode(stream).await?;
+		debug!("client: got response to request {} (path {})", id, path);
+		Resp::from_enc(resp_enc)
 	}
 }
 
@@ -235,14 +222,71 @@ impl SendLoop for ClientConn {}
 
 #[async_trait]
 impl RecvLoop for ClientConn {
-	fn recv_handler(self: &Arc<Self>, id: RequestID, msg: Vec<u8>) {
-		trace!("ClientConn recv_handler {} ({} bytes)", id, msg.len());
+	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(msg).is_err() {
+			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: 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()
+	}
+}
diff --git a/src/endpoint.rs b/src/endpoint.rs
index 42e9a98..3cafafe 100644
--- a/src/endpoint.rs
+++ b/src/endpoint.rs
@@ -1,33 +1,25 @@
-use std::borrow::Borrow;
 use std::marker::PhantomData;
 use std::sync::Arc;
 
 use arc_swap::ArcSwapOption;
 use async_trait::async_trait;
 
-use serde::{Deserialize, Serialize};
-
 use crate::error::Error;
+use crate::message::*;
 use crate::netapp::*;
-use crate::proto::*;
-use crate::util::*;
-
-/// This trait should be implemented by all messages your application
-/// wants to handle
-pub trait Message: Serialize + for<'de> Deserialize<'de> + Send + Sync {
-	type Response: Serialize + for<'de> Deserialize<'de> + Send + Sync;
-}
 
 /// This trait should be implemented by an object of your application
-/// that can handle a message of type `M`.
+/// 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 EndpointHandler<M>: Send + Sync
+pub trait StreamingEndpointHandler<M>: Send + Sync
 where
 	M: Message,
 {
-	async fn handle(self: &Arc<Self>, m: &M, from: NodeID) -> M::Response;
+	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,
@@ -35,12 +27,41 @@ where
 /// use the unit type `()` as the handler type:
 /// it will panic if it is ever made to handle request.
 #[async_trait]
-impl<M: Message + 'static> EndpointHandler<M> for () {
+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`.
@@ -50,13 +71,13 @@ impl<M: Message + 'static> EndpointHandler<M> for () {
 /// 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 `EndpointHandler`).
+/// endpoint messages (see `StreamingEndpointHandler`).
 pub struct Endpoint<M, H>
 where
 	M: Message,
-	H: EndpointHandler<M>,
+	H: StreamingEndpointHandler<M>,
 {
-	phantom: PhantomData<M>,
+	_phantom: PhantomData<M>,
 	netapp: Arc<NetApp>,
 	path: String,
 	handler: ArcSwapOption<H>,
@@ -65,11 +86,11 @@ where
 impl<M, H> Endpoint<M, H>
 where
 	M: Message,
-	H: EndpointHandler<M>,
+	H: StreamingEndpointHandler<M>,
 {
 	pub(crate) fn new(netapp: Arc<NetApp>, path: String) -> Self {
 		Self {
-			phantom: PhantomData::default(),
+			_phantom: PhantomData::default(),
 			netapp,
 			path,
 			handler: ArcSwapOption::from(None),
@@ -88,20 +109,22 @@ where
 	}
 
 	/// Call this endpoint on a remote node (or on the local node,
-	/// for that matter)
-	pub async fn call<B>(
+	/// 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: B,
+		req: T,
 		prio: RequestPriority,
-	) -> Result<<M as Message>::Response, Error>
+	) -> Result<Resp<M>, Error>
 	where
-		B: Borrow<M>,
+		T: IntoReq<M>,
 	{
 		if *target == self.netapp.id {
 			match self.handler.load_full() {
 				None => Err(Error::NoHandler),
-				Some(h) => Ok(h.handle(req.borrow(), self.netapp.id).await),
+				Some(h) => Ok(h.handle(req.into_req_local(), self.netapp.id).await),
 			}
 		} else {
 			let conn = self
@@ -116,10 +139,22 @@ where
 					"Not connected: {}",
 					hex::encode(&target[..8])
 				))),
-				Some(c) => c.call(req, self.path.as_str(), prio).await,
+				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 ----
@@ -128,7 +163,7 @@ pub(crate) type DynEndpoint = Box<dyn GenericEndpoint + Send + Sync>;
 
 #[async_trait]
 pub(crate) trait GenericEndpoint {
-	async fn handle(&self, buf: &[u8], from: NodeID) -> Result<Vec<u8>, Error>;
+	async fn handle(&self, req_enc: ReqEnc, from: NodeID) -> Result<RespEnc, Error>;
 	fn drop_handler(&self);
 	fn clone_endpoint(&self) -> DynEndpoint;
 }
@@ -137,22 +172,21 @@ pub(crate) trait GenericEndpoint {
 pub(crate) struct EndpointArc<M, H>(pub(crate) Arc<Endpoint<M, H>>)
 where
 	M: Message,
-	H: EndpointHandler<M>;
+	H: StreamingEndpointHandler<M>;
 
 #[async_trait]
 impl<M, H> GenericEndpoint for EndpointArc<M, H>
 where
-	M: Message + 'static,
-	H: EndpointHandler<M> + 'static,
+	M: Message,
+	H: StreamingEndpointHandler<M> + 'static,
 {
-	async fn handle(&self, buf: &[u8], from: NodeID) -> Result<Vec<u8>, Error> {
+	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 = rmp_serde::decode::from_read_ref::<_, M>(buf)?;
-				let res = h.handle(&req, from).await;
-				let res_bytes = rmp_to_vec_all_named(&res)?;
-				Ok(res_bytes)
+				let req = Req::from_enc(req_enc)?;
+				let res = h.handle(req, from).await;
+				Ok(res.into_enc()?)
 			}
 		}
 	}
diff --git a/src/error.rs b/src/error.rs
index 99acdd1..c0aeeac 100644
--- a/src/error.rs
+++ b/src/error.rs
@@ -1,6 +1,6 @@
-use err_derive::Error;
 use std::io;
 
+use err_derive::Error;
 use log::error;
 
 #[derive(Debug, Error)]
@@ -25,6 +25,15 @@ pub enum 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),
 
@@ -36,28 +45,6 @@ pub enum Error {
 
 	#[error(display = "Version mismatch: {}", _0)]
 	VersionMismatch(String),
-
-	#[error(display = "Remote error {}: {}", _0, _1)]
-	Remote(u8, String),
-}
-
-impl Error {
-	pub fn code(&self) -> u8 {
-		match self {
-			Self::Io(_) => 100,
-			Self::TokioJoin(_) => 110,
-			Self::OneshotRecv(_) => 111,
-			Self::RMPEncode(_) => 10,
-			Self::RMPDecode(_) => 11,
-			Self::UTF8(_) => 12,
-			Self::NoHandler => 20,
-			Self::ConnectionClosed => 21,
-			Self::Handshake(_) => 30,
-			Self::VersionMismatch(_) => 31,
-			Self::Remote(c, _) => *c,
-			Self::Message(_) => 99,
-		}
-	}
 }
 
 impl<T> From<tokio::sync::watch::error::SendError<T>> for Error {
@@ -101,3 +88,39 @@ where
 		}
 	}
 }
+
+// ---- 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,
+	}
+}
diff --git a/src/lib.rs b/src/lib.rs
index cb24337..8e30e40 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -13,21 +13,23 @@
 //! about message priorization.
 //! Also check out the examples to learn how to use this crate.
 
+pub mod bytes_buf;
 pub mod error;
+pub mod stream;
 pub mod util;
 
 pub mod endpoint;
-pub mod proto;
+pub mod message;
 
 mod client;
-mod proto2;
+mod recv;
+mod send;
 mod server;
 
 pub mod netapp;
 pub mod peering;
 
 pub use crate::netapp::*;
-pub use util::{NetworkKey, NodeID, NodeKey};
 
 #[cfg(test)]
 mod test;
diff --git a/src/message.rs b/src/message.rs
new file mode 100644
index 0000000..9cc1a3e
--- /dev/null
+++ b/src/message.rs
@@ -0,0 +1,522 @@
+use std::fmt;
+use std::marker::PhantomData;
+use std::sync::Arc;
+
+use bytes::{BufMut, Bytes, BytesMut};
+use rand::prelude::*;
+use serde::{Deserialize, Serialize};
+
+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. 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;
+}
+
+// ----
+
+/// 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<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,
+	}
+}
diff --git a/src/netapp.rs b/src/netapp.rs
index e9efa2e..b1ad9db 100644
--- a/src/netapp.rs
+++ b/src/netapp.rs
@@ -20,9 +20,15 @@ use tokio::sync::{mpsc, watch};
 use crate::client::*;
 use crate::endpoint::*;
 use crate::error::*;
-use crate::proto::*;
+use crate::message::*;
 use crate::server::*;
-use crate::util::*;
+
+/// 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,
@@ -30,7 +36,7 @@ use crate::util::*;
 pub(crate) type VersionTag = [u8; 16];
 
 /// Value of the Netapp version used in the version tag
-pub(crate) const NETAPP_VERSION_TAG: u64 = 0x6e65746170700004; // netapp 0x0004
+pub(crate) const NETAPP_VERSION_TAG: u64 = 0x6e65746170700005; // netapp 0x0005
 
 #[derive(Serialize, Deserialize, Debug)]
 pub(crate) struct HelloMessage {
@@ -152,7 +158,7 @@ impl NetApp {
 	pub fn endpoint<M, H>(self: &Arc<Self>, path: String) -> Arc<Endpoint<M, H>>
 	where
 		M: Message + 'static,
-		H: EndpointHandler<M> + 'static,
+		H: StreamingEndpointHandler<M> + 'static,
 	{
 		let endpoint = Arc::new(Endpoint::<M, H>::new(self.clone(), path.clone()));
 		let endpoint_arc = EndpointArc(endpoint.clone());
@@ -397,13 +403,14 @@ impl NetApp {
 				hello_endpoint
 					.call(
 						&conn.peer_id,
-						&HelloMessage {
+						HelloMessage {
 							server_addr,
 							server_port,
 						},
 						PRIO_NORMAL,
 					)
 					.await
+					.map(|_| ())
 					.log_err("Sending hello message");
 			});
 		}
diff --git a/src/peering/basalt.rs b/src/peering/basalt.rs
index cd70c2b..2c27dba 100644
--- a/src/peering/basalt.rs
+++ b/src/peering/basalt.rs
@@ -14,8 +14,8 @@ 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 --
@@ -331,7 +331,7 @@ impl Basalt {
 	async fn do_pull(self: Arc<Self>, peer: NodeID) {
 		match self
 			.pull_endpoint
-			.call(&peer, &PullMessage {}, PRIO_NORMAL)
+			.call(&peer, PullMessage {}, PRIO_NORMAL)
 			.await
 		{
 			Ok(resp) => {
@@ -346,7 +346,7 @@ impl Basalt {
 
 	async fn do_push(self: Arc<Self>, peer: NodeID) {
 		let push_msg = self.make_push_message();
-		match self.push_endpoint.call(&peer, &push_msg, PRIO_NORMAL).await {
+		match self.push_endpoint.call(&peer, push_msg, PRIO_NORMAL).await {
 			Ok(_) => {
 				trace!("KYEV PEXo {}", hex::encode(peer));
 			}
diff --git a/src/peering/fullmesh.rs b/src/peering/fullmesh.rs
index 208cfe4..fb2e3d1 100644
--- a/src/peering/fullmesh.rs
+++ b/src/peering/fullmesh.rs
@@ -17,7 +17,8 @@ use sodiumoxide::crypto::hash;
 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);
@@ -80,6 +81,7 @@ impl PeerInfoInternal {
 	}
 }
 
+/// Information that the full mesh peering strategy can return about the peers it knows of
 #[derive(Copy, Clone, Debug)]
 pub struct PeerInfo {
 	/// The node's identifier (its public key)
@@ -382,7 +384,7 @@ impl FullMeshPeeringStrategy {
 			ping_time
 		);
 		let ping_response = select! {
-			r = self.ping_endpoint.call(&id, &ping_msg, PRIO_HIGH) => r,
+			r = self.ping_endpoint.call(&id, ping_msg, PRIO_HIGH) => r,
 			_ = tokio::time::sleep(PING_TIMEOUT) => Err(Error::Message("Ping timeout".into())),
 		};
 
@@ -434,7 +436,7 @@ impl FullMeshPeeringStrategy {
 		let pex_message = PeerListMessage { list: peer_list };
 		match self
 			.peer_list_endpoint
-			.call(id, &pex_message, PRIO_BACKGROUND)
+			.call(id, pex_message, PRIO_BACKGROUND)
 			.await
 		{
 			Err(e) => warn!("Error doing peer exchange: {}", e),
diff --git a/src/proto.rs b/src/proto.rs
deleted file mode 100644
index 8f7e70f..0000000
--- a/src/proto.rs
+++ /dev/null
@@ -1,358 +0,0 @@
-use std::collections::{HashMap, VecDeque};
-use std::fmt::Write;
-use std::sync::Arc;
-
-use log::trace;
-
-use futures::{AsyncReadExt, AsyncWriteExt};
-use kuska_handshake::async_std::BoxStreamWrite;
-
-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;
-
-// Messages are sent by chunks
-// Chunk format:
-// - u32 BE: request id (same for request and response)
-// - u16 BE: chunk length, possibly with CHUNK_HAS_CONTINUATION flag
-//					when this is not the last chunk of the message
-// - [u8; chunk_length] chunk data
-
-pub(crate) type RequestID = u32;
-type ChunkLength = u16;
-const MAX_CHUNK_LENGTH: ChunkLength = 0x4000;
-const CHUNK_HAS_CONTINUATION: ChunkLength = 0x8000;
-
-struct SendQueueItem {
-	id: RequestID,
-	prio: RequestPriority,
-	data: Vec<u8>,
-	cursor: usize,
-}
-
-struct SendQueue {
-	items: VecDeque<(u8, VecDeque<SendQueueItem>)>,
-}
-
-impl SendQueue {
-	fn new() -> Self {
-		Self {
-			items: VecDeque::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, VecDeque::new()));
-				i
-			}
-		};
-		self.items[pos_prio].1.push_back(item);
-	}
-	fn pop(&mut self) -> Option<SendQueueItem> {
-		match self.items.pop_front() {
-			None => None,
-			Some((prio, mut items_at_prio)) => {
-				let ret = items_at_prio.pop_front();
-				if !items_at_prio.is_empty() {
-					self.items.push_front((prio, items_at_prio));
-				}
-				ret.or_else(|| self.pop())
-			}
-		}
-	}
-	fn is_empty(&self) -> bool {
-		self.items.iter().all(|(_k, v)| v.is_empty())
-	}
-	fn dump(&self) -> String {
-		let mut ret = String::new();
-		for (prio, q) in self.items.iter() {
-			for item in q.iter() {
-				write!(
-					&mut ret,
-					" [{} {} ({})]",
-					prio,
-					item.data.len() - item.cursor,
-					item.id
-				)
-				.unwrap();
-			}
-		}
-		ret
-	}
-}
-
-/// 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>,
-		mut msg_recv: mpsc::UnboundedReceiver<(RequestID, RequestPriority, Vec<u8>)>,
-		mut write: BoxStreamWrite<W>,
-		debug_name: String,
-	) -> Result<(), Error>
-	where
-		W: AsyncWriteExt + Unpin + Send + Sync,
-	{
-		let mut sending = SendQueue::new();
-		let mut should_exit = false;
-		while !should_exit || !sending.is_empty() {
-			trace!("send_loop({}): queue = {}", debug_name, sending.dump());
-			if let Ok((id, prio, data)) = msg_recv.try_recv() {
-				trace!(
-					"send_loop({}): new message to send, id = {}, prio = {}, {} bytes",
-					debug_name,
-					id,
-					prio,
-					data.len()
-				);
-				sending.push(SendQueueItem {
-					id,
-					prio,
-					data,
-					cursor: 0,
-				});
-			} else if let Some(mut item) = sending.pop() {
-				trace!(
-					"send_loop({}): sending bytes for {} ({} bytes, {} already sent)",
-					debug_name,
-					item.id,
-					item.data.len(),
-					item.cursor
-				);
-				let header_id = RequestID::to_be_bytes(item.id);
-				write.write_all(&header_id[..]).await?;
-
-				if item.data.len() - item.cursor > MAX_CHUNK_LENGTH as usize {
-					let size_header =
-						ChunkLength::to_be_bytes(MAX_CHUNK_LENGTH | CHUNK_HAS_CONTINUATION);
-					write.write_all(&size_header[..]).await?;
-
-					let new_cursor = item.cursor + MAX_CHUNK_LENGTH 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 ChunkLength;
-
-					let size_header = ChunkLength::to_be_bytes(send_len);
-					write.write_all(&size_header[..]).await?;
-
-					write.write_all(&item.data[item.cursor..]).await?;
-				}
-				write.flush().await?;
-			} else {
-				let sth = msg_recv.recv().await;
-				if let Some((id, prio, data)) = sth {
-					trace!(
-						"send_loop({}): new message to send, id = {}, prio = {}, {} bytes",
-						debug_name,
-						id,
-						prio,
-						data.len()
-					);
-					sending.push(SendQueueItem {
-						id,
-						prio,
-						data,
-						cursor: 0,
-					});
-				} else {
-					should_exit = true;
-				}
-			}
-		}
-
-		let _ = write.goodbye().await;
-		Ok(())
-	}
-}
-
-/// 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, msg: Vec<u8>);
-
-	async fn recv_loop<R>(self: Arc<Self>, mut read: R, debug_name: String) -> Result<(), Error>
-	where
-		R: AsyncReadExt + Unpin + Send + Sync,
-	{
-		let mut receiving = HashMap::new();
-		loop {
-			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);
-			trace!(
-				"recv_loop({}): got header id = {}, size = 0x{:04x} ({} bytes)",
-				debug_name,
-				id,
-				size,
-				size & !CHUNK_HAS_CONTINUATION
-			);
-
-			let has_cont = (size & CHUNK_HAS_CONTINUATION) != 0;
-			let size = size & !CHUNK_HAS_CONTINUATION;
-
-			let mut next_slice = vec![0; size as usize];
-			read.read_exact(&mut next_slice[..]).await?;
-			trace!("recv_loop({}): read {} bytes", debug_name, 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 {
-				self.recv_handler(id, msg_bytes);
-			}
-		}
-		Ok(())
-	}
-}
-
-#[cfg(test)]
-mod test {
-	use super::*;
-
-	#[test]
-	fn test_priority_queue() {
-		let i1 = SendQueueItem {
-			id: 1,
-			prio: PRIO_NORMAL,
-			data: vec![],
-			cursor: 0,
-		};
-		let i2 = SendQueueItem {
-			id: 2,
-			prio: PRIO_HIGH,
-			data: vec![],
-			cursor: 0,
-		};
-		let i2bis = SendQueueItem {
-			id: 20,
-			prio: PRIO_HIGH,
-			data: vec![],
-			cursor: 0,
-		};
-		let i3 = SendQueueItem {
-			id: 3,
-			prio: PRIO_HIGH | PRIO_SECONDARY,
-			data: vec![],
-			cursor: 0,
-		};
-		let i4 = SendQueueItem {
-			id: 4,
-			prio: PRIO_BACKGROUND | PRIO_SECONDARY,
-			data: vec![],
-			cursor: 0,
-		};
-		let i5 = SendQueueItem {
-			id: 5,
-			prio: PRIO_BACKGROUND | PRIO_PRIMARY,
-			data: vec![],
-			cursor: 0,
-		};
-
-		let mut q = SendQueue::new();
-
-		q.push(i1); // 1
-		let a = q.pop().unwrap(); // empty -> 1
-		assert_eq!(a.id, 1);
-		assert!(q.pop().is_none());
-
-		q.push(a); // 1
-		q.push(i2); // 2 1
-		q.push(i2bis); // [2 20] 1
-		let a = q.pop().unwrap(); // 20 1 -> 2
-		assert_eq!(a.id, 2);
-		let b = q.pop().unwrap(); // 1 -> 20
-		assert_eq!(b.id, 20);
-		let c = q.pop().unwrap(); // empty -> 1
-		assert_eq!(c.id, 1);
-		assert!(q.pop().is_none());
-
-		q.push(a); // 2
-		q.push(b); // [2 20]
-		q.push(c); // [2 20] 1
-		q.push(i3); // [2 20] 3 1
-		q.push(i4); // [2 20] 3 1 4
-		q.push(i5); // [2 20] 3 1 5 4
-
-		let a = q.pop().unwrap(); // 20 3 1 5 4 -> 2
-		assert_eq!(a.id, 2);
-		q.push(a); // [20 2] 3 1 5 4
-
-		let a = q.pop().unwrap(); // 2 3 1 5 4 -> 20
-		assert_eq!(a.id, 20);
-		let b = q.pop().unwrap(); // 3 1 5 4 -> 2
-		assert_eq!(b.id, 2);
-		q.push(b); // 2 3 1 5 4
-		let b = q.pop().unwrap(); // 3 1 5 4 -> 2
-		assert_eq!(b.id, 2);
-		let c = q.pop().unwrap(); // 1 5 4 -> 3
-		assert_eq!(c.id, 3);
-		q.push(b); // 2 1 5 4
-		let b = q.pop().unwrap(); // 1 5 4 -> 2
-		assert_eq!(b.id, 2);
-		let e = q.pop().unwrap(); // 5 4 -> 1
-		assert_eq!(e.id, 1);
-		let f = q.pop().unwrap(); // 4 -> 5
-		assert_eq!(f.id, 5);
-		let g = q.pop().unwrap(); // empty -> 4
-		assert_eq!(g.id, 4);
-		assert!(q.pop().is_none());
-	}
-}
diff --git a/src/proto2.rs b/src/proto2.rs
deleted file mode 100644
index 7210781..0000000
--- a/src/proto2.rs
+++ /dev/null
@@ -1,75 +0,0 @@
-use crate::error::*;
-use crate::proto::*;
-
-pub(crate) struct QueryMessage<'a> {
-	pub(crate) prio: RequestPriority,
-	pub(crate) path: &'a [u8],
-	pub(crate) telemetry_id: Option<Vec<u8>>,
-	pub(crate) body: &'a [u8],
-}
-
-/// QueryMessage encoding:
-/// - priority: u8
-/// - path length: u8
-/// - path: [u8; path length]
-/// - telemetry id length: u8
-/// - telemetry id: [u8; telemetry id length]
-/// - body [u8; ..]
-impl<'a> QueryMessage<'a> {
-	pub(crate) fn encode(self) -> Vec<u8> {
-		let tel_len = match &self.telemetry_id {
-			Some(t) => t.len(),
-			None => 0,
-		};
-
-		let mut ret = Vec::with_capacity(10 + self.path.len() + tel_len + self.body.len());
-
-		ret.push(self.prio);
-
-		ret.push(self.path.len() as u8);
-		ret.extend_from_slice(self.path);
-
-		if let Some(t) = self.telemetry_id {
-			ret.push(t.len() as u8);
-			ret.extend(t);
-		} else {
-			ret.push(0u8);
-		}
-
-		ret.extend_from_slice(self.body);
-
-		ret
-	}
-
-	pub(crate) fn decode(bytes: &'a [u8]) -> Result<Self, Error> {
-		if bytes.len() < 3 {
-			return Err(Error::Message("Invalid protocol message".into()));
-		}
-
-		let path_length = bytes[1] as usize;
-		if bytes.len() < 3 + path_length {
-			return Err(Error::Message("Invalid protocol message".into()));
-		}
-
-		let telemetry_id_len = bytes[2 + path_length] as usize;
-		if bytes.len() < 3 + path_length + telemetry_id_len {
-			return Err(Error::Message("Invalid protocol message".into()));
-		}
-
-		let path = &bytes[2..2 + path_length];
-		let telemetry_id = if telemetry_id_len > 0 {
-			Some(bytes[3 + path_length..3 + path_length + telemetry_id_len].to_vec())
-		} else {
-			None
-		};
-
-		let body = &bytes[3 + path_length + telemetry_id_len..];
-
-		Ok(Self {
-			prio: bytes[0],
-			path,
-			telemetry_id,
-			body,
-		})
-	}
-}
diff --git a/src/recv.rs b/src/recv.rs
new file mode 100644
index 0000000..0de7bef
--- /dev/null
+++ b/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(())
+	}
+}
diff --git a/src/send.rs b/src/send.rs
new file mode 100644
index 0000000..0db0ba7
--- /dev/null
+++ b/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(())
+	}
+}
diff --git a/src/server.rs b/src/server.rs
index a835959..cd367c4 100644
--- a/src/server.rs
+++ b/src/server.rs
@@ -1,9 +1,17 @@
+use std::collections::HashMap;
 use std::net::SocketAddr;
-use std::sync::Arc;
+use std::sync::{Arc, Mutex};
 
 use arc_swap::ArcSwapOption;
-use bytes::Bytes;
-use log::{debug, trace};
+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::{
@@ -15,21 +23,12 @@ use opentelemetry_contrib::trace::propagator::binary::*;
 #[cfg(feature = "telemetry")]
 use rand::{thread_rng, Rng};
 
-use tokio::net::TcpStream;
-use tokio::select;
-use tokio::sync::{mpsc, watch};
-use tokio_util::compat::*;
-
-use futures::io::{AsyncReadExt, AsyncWriteExt};
-
-use async_trait::async_trait;
-
-use kuska_handshake::async_std::{handshake_server, BoxStream};
-
 use crate::error::*;
+use crate::message::*;
 use crate::netapp::*;
-use crate::proto::*;
-use crate::proto2::*;
+use crate::recv::*;
+use crate::send::*;
+use crate::stream::*;
 use crate::util::*;
 
 // The client and server connection structs (client.rs and server.rs)
@@ -55,7 +54,8 @@ pub(crate) struct ServerConn {
 
 	netapp: Arc<NetApp>,
 
-	resp_send: ArcSwapOption<mpsc::UnboundedSender<(RequestID, RequestPriority, Vec<u8>)>>,
+	resp_send: ArcSwapOption<mpsc::UnboundedSender<SendItem>>,
+	running_handlers: Mutex<HashMap<RequestID, tokio::task::JoinHandle<()>>>,
 }
 
 impl ServerConn {
@@ -101,6 +101,7 @@ impl ServerConn {
 			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());
@@ -126,9 +127,8 @@ impl ServerConn {
 		Ok(())
 	}
 
-	async fn recv_handler_aux(self: &Arc<Self>, bytes: &[u8]) -> Result<Vec<u8>, Error> {
-		let msg = QueryMessage::decode(bytes)?;
-		let path = String::from_utf8(msg.path.to_vec())?;
+	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();
@@ -140,9 +140,9 @@ impl ServerConn {
 				if #[cfg(feature = "telemetry")] {
 					let tracer = opentelemetry::global::tracer("netapp");
 
-					let mut span = if let Some(telemetry_id) = msg.telemetry_id {
+					let mut span = if !req_enc.telemetry_id.is_empty() {
 						let propagator = BinaryPropagator::new();
-						let context = propagator.from_bytes(telemetry_id);
+						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)
@@ -157,13 +157,13 @@ impl ServerConn {
 							.start(&tracer)
 					};
 					span.set_attribute(KeyValue::new("path", path.to_string()));
-					span.set_attribute(KeyValue::new("len_query", msg.body.len() as i64));
+					span.set_attribute(KeyValue::new("len_query_msg", req_enc.msg.len() as i64));
 
-					handler.handle(msg.body, self.peer_id)
+					handler.handle(req_enc, self.peer_id)
 						.with_context(Context::current_with_span(span))
 						.await
 				} else {
-					handler.handle(msg.body, self.peer_id).await
+					handler.handle(req_enc, self.peer_id).await
 				}
 			}
 		} else {
@@ -176,35 +176,47 @@ impl SendLoop for ServerConn {}
 
 #[async_trait]
 impl RecvLoop for ServerConn {
-	fn recv_handler(self: &Arc<Self>, id: RequestID, bytes: Vec<u8>) {
-		let resp_send = self.resp_send.load_full().unwrap();
+	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();
-		tokio::spawn(async move {
-			trace!("ServerConn recv_handler {} ({} bytes)", id, bytes.len());
-			let bytes: Bytes = bytes.into();
-
-			let prio = if !bytes.is_empty() { bytes[0] } else { 0u8 };
-			let resp = self2.recv_handler_aux(&bytes[..]).await;
-
-			let resp_bytes = match resp {
-				Ok(rb) => {
-					let mut resp_bytes = vec![0u8];
-					resp_bytes.extend(rb);
-					resp_bytes
-				}
-				Err(e) => {
-					let mut resp_bytes = vec![e.code()];
-					resp_bytes.extend(e.to_string().into_bytes());
-					resp_bytes
-				}
+		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)),
 			};
 
-			trace!("ServerConn sending response to {}: ", id);
+			debug!("server: sending response to {}", id);
 
+			let (resp_stream, resp_order) = RespEnc::encode(resp_enc_result);
 			resp_send
-				.send((id, prio, resp_bytes))
-				.log_err("ServerConn recv_handler send resp");
+				.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));
+		}
 	}
 }
diff --git a/src/stream.rs b/src/stream.rs
new file mode 100644
index 0000000..88c3fed
--- /dev/null
+++ b/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)
+}
diff --git a/src/test.rs b/src/test.rs
index 82c7ba6..19759cc 100644
--- a/src/test.rs
+++ b/src/test.rs
@@ -14,6 +14,7 @@ use crate::NodeID;
 
 #[tokio::test(flavor = "current_thread")]
 async fn test_with_basic_scheduler() {
+	env_logger::init();
 	run_test().await
 }
 
diff --git a/src/util.rs b/src/util.rs
index f4dfac7..425d26f 100644
--- a/src/util.rs
+++ b/src/util.rs
@@ -1,18 +1,12 @@
 use std::net::SocketAddr;
 use std::net::ToSocketAddrs;
 
-use serde::Serialize;
-
 use log::info;
+use serde::Serialize;
 
 use tokio::sync::watch;
 
-/// 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;
+use crate::netapp::*;
 
 /// Utility function: encodes any serializable value in MessagePack binary format
 /// using the RMP library.