Compare commits

...

8 Commits

8 changed files with 121 additions and 19 deletions

View File

@@ -1,5 +1,28 @@
# Changelog
## 2026-03-16 - 4.6.0 - feat(remoteingress-core)
add adaptive per-stream flow control based on active stream counts
- Track active stream counts on edge and hub connections to size per-stream flow control windows dynamically.
- Cap WINDOW_UPDATE increments and read sizes to the adaptive window so bandwidth is shared more evenly across concurrent streams.
- Apply the adaptive logic to both upload and download paths on edge and hub stream handlers.
## 2026-03-16 - 4.5.12 - fix(remoteingress-core)
improve tunnel liveness handling and enable TCP keepalive for accepted client sockets
- Avoid disconnecting edges when PING or PONG frames cannot be queued because the control channel is temporarily full.
- Enable TCP_NODELAY and TCP keepalive on accepted client connections to help detect stale or dropped clients.
## 2026-03-16 - 4.5.11 - fix(repo)
no changes to commit
## 2026-03-16 - 4.5.10 - fix(remoteingress-core)
guard zero-window reads to avoid false EOF handling on stalled streams
- Prevent upload and download loops from calling read on an empty buffer when flow-control window remains at 0 after stall timeout
- Log a warning and close the affected stream instead of misinterpreting Ok(0) as end-of-file
## 2026-03-16 - 4.5.9 - fix(remoteingress-core)
delay stream close until downstream response draining finishes to prevent truncated transfers

View File

@@ -1,6 +1,6 @@
{
"name": "@serve.zone/remoteingress",
"version": "4.5.9",
"version": "4.6.0",
"private": false,
"description": "Edge ingress tunnel for DcRouter - accepts incoming TCP connections at network edge and tunnels them to DcRouter SmartProxy preserving client IP via PROXY protocol v1.",
"main": "dist_ts/index.js",

13
rust/Cargo.lock generated
View File

@@ -558,6 +558,7 @@ dependencies = [
"rustls-pemfile",
"serde",
"serde_json",
"socket2 0.5.10",
"tokio",
"tokio-rustls",
"tokio-util",
@@ -701,6 +702,16 @@ version = "1.15.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "67b1b7a3b5fe4f1376887184045fcf45c69e92af734b7aaddc05fb777b6fbd03"
[[package]]
name = "socket2"
version = "0.5.10"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "e22376abed350d73dd1cd119b57ffccad95b4e585a7cda43e286245ce23c0678"
dependencies = [
"libc",
"windows-sys 0.52.0",
]
[[package]]
name = "socket2"
version = "0.6.2"
@@ -765,7 +776,7 @@ dependencies = [
"parking_lot",
"pin-project-lite",
"signal-hook-registry",
"socket2",
"socket2 0.6.2",
"tokio-macros",
"windows-sys 0.61.2",
]

View File

@@ -14,3 +14,4 @@ serde_json = "1"
log = "0.4"
rustls-pemfile = "2"
tokio-util = "0.7"
socket2 = "0.5"

View File

@@ -494,8 +494,10 @@ async fn connect_to_hub_and_run(
FRAME_PING => {
let pong_frame = encode_frame(0, FRAME_PONG, &[]);
if tunnel_writer_tx.try_send(pong_frame).is_err() {
log::warn!("Failed to send PONG, writer channel full/closed");
break EdgeLoopResult::Reconnect;
// Control channel full (WINDOW_UPDATE burst from many streams).
// DON'T disconnect — the 45s liveness timeout gives margin
// for the channel to drain and the next PONG to succeed.
log::warn!("PONG send failed, control channel full — skipping this cycle");
}
log::trace!("Received PING from hub, sent PONG");
}
@@ -588,6 +590,15 @@ fn apply_port_config(
accept_result = listener.accept() => {
match accept_result {
Ok((client_stream, client_addr)) => {
// TCP keepalive detects dead clients that disappear without FIN.
// Without this, zombie streams accumulate and never get cleaned up.
let _ = client_stream.set_nodelay(true);
let ka = socket2::TcpKeepalive::new()
.with_time(Duration::from_secs(60));
#[cfg(target_os = "linux")]
let ka = ka.with_interval(Duration::from_secs(60));
let _ = socket2::SockRef::from(&client_stream).set_tcp_keepalive(&ka);
let stream_id = next_stream_id.fetch_add(1, Ordering::Relaxed);
let tunnel_ctrl_tx = tunnel_ctrl_tx.clone();
let tunnel_data_tx = tunnel_data_tx.clone();
@@ -609,6 +620,7 @@ fn apply_port_config(
tunnel_data_tx,
client_writers,
client_token,
Arc::clone(&active_streams),
)
.await;
active_streams.fetch_sub(1, Ordering::Relaxed);
@@ -640,6 +652,7 @@ async fn handle_client_connection(
tunnel_data_tx: mpsc::Sender<Vec<u8>>,
client_writers: Arc<Mutex<HashMap<u32, EdgeStreamState>>>,
client_token: CancellationToken,
active_streams: Arc<AtomicU32>,
) {
let client_ip = client_addr.ip().to_string();
let client_port = client_addr.port();
@@ -673,6 +686,7 @@ async fn handle_client_connection(
// After writing to client TCP, send WINDOW_UPDATE to hub so it can send more
let hub_to_client_token = client_token.clone();
let wu_tx = tunnel_ctrl_tx.clone();
let active_streams_h2c = Arc::clone(&active_streams);
let mut hub_to_client = tokio::spawn(async move {
let mut consumed_since_update: u32 = 0;
loop {
@@ -684,12 +698,20 @@ async fn handle_client_connection(
if client_write.write_all(&data).await.is_err() {
break;
}
// Track consumption for flow control
// Track consumption for adaptive flow control.
// The increment is capped to the adaptive window so the sender's
// effective window shrinks to match current demand (fewer streams
// = larger window, more streams = smaller window per stream).
consumed_since_update += len;
if consumed_since_update >= WINDOW_UPDATE_THRESHOLD {
let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE, consumed_since_update);
let adaptive_window = remoteingress_protocol::compute_window_for_stream_count(
active_streams_h2c.load(Ordering::Relaxed),
);
let threshold = adaptive_window / 2;
if consumed_since_update >= threshold {
let increment = consumed_since_update.min(adaptive_window);
let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE, increment);
if wu_tx.try_send(frame).is_ok() {
consumed_since_update = 0;
consumed_since_update -= increment;
}
// If try_send fails, keep accumulating — retry on next threshold
}
@@ -726,9 +748,20 @@ async fn handle_client_connection(
}
if client_token.is_cancelled() { break; }
// Limit read size to available window
// Limit read size to available window.
// IMPORTANT: if window is 0 (stall timeout fired), we must NOT
// read into an empty buffer — read(&mut buf[..0]) returns Ok(0)
// which would be falsely interpreted as EOF.
let w = send_window.load(Ordering::Acquire) as usize;
let max_read = w.min(buf.len());
if w == 0 {
log::warn!("Stream {} upload: window still 0 after stall timeout, closing", stream_id);
break;
}
// Adaptive: cap read to current per-stream target window
let adaptive_cap = remoteingress_protocol::compute_window_for_stream_count(
active_streams.load(Ordering::Relaxed),
) as usize;
let max_read = w.min(buf.len()).min(adaptive_cap);
tokio::select! {
read_result = client_read.read(&mut buf[..max_read]) => {

View File

@@ -373,6 +373,9 @@ async fn handle_edge_connection(
);
}
// Per-edge active stream counter for adaptive flow control
let edge_stream_count = Arc::new(AtomicU32::new(0));
// QoS dual-channel tunnel writer: control frames (PING/PONG/WINDOW_UPDATE/CLOSE)
// have priority over data frames (DATA_BACK). This prevents PING starvation under load.
let (ctrl_tx, mut ctrl_rx) = mpsc::channel::<Vec<u8>>(64);
@@ -509,8 +512,10 @@ async fn handle_edge_connection(
}
// Spawn task: connect to SmartProxy, send PROXY header, pipe data
let stream_counter = Arc::clone(&edge_stream_count);
tokio::spawn(async move {
let _permit = permit; // hold semaphore permit until stream completes
stream_counter.fetch_add(1, Ordering::Relaxed);
let result = async {
// A2: Connect to SmartProxy with timeout
@@ -533,6 +538,7 @@ async fn handle_edge_connection(
// After writing to upstream, send WINDOW_UPDATE_BACK to edge
let writer_token = stream_token.clone();
let wub_tx = writer_tx.clone();
let stream_counter_w = Arc::clone(&stream_counter);
let writer_for_edge_data = tokio::spawn(async move {
let mut consumed_since_update: u32 = 0;
loop {
@@ -558,12 +564,18 @@ async fn handle_edge_connection(
break;
}
}
// Track consumption for flow control
// Track consumption for adaptive flow control.
// Increment capped to adaptive window to limit per-stream in-flight data.
consumed_since_update += len;
if consumed_since_update >= WINDOW_UPDATE_THRESHOLD {
let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE_BACK, consumed_since_update);
let adaptive_window = remoteingress_protocol::compute_window_for_stream_count(
stream_counter_w.load(Ordering::Relaxed),
);
let threshold = adaptive_window / 2;
if consumed_since_update >= threshold {
let increment = consumed_since_update.min(adaptive_window);
let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE_BACK, increment);
if wub_tx.try_send(frame).is_ok() {
consumed_since_update = 0;
consumed_since_update -= increment;
}
// If try_send fails, keep accumulating — retry on next threshold
}
@@ -601,9 +613,20 @@ async fn handle_edge_connection(
}
if stream_token.is_cancelled() { break; }
// Limit read size to available window
// Limit read size to available window.
// IMPORTANT: if window is 0 (stall timeout fired), we must NOT
// read into an empty buffer — read(&mut buf[..0]) returns Ok(0)
// which would be falsely interpreted as EOF.
let w = send_window.load(Ordering::Acquire) as usize;
let max_read = w.min(buf.len());
if w == 0 {
log::warn!("Stream {} download: window still 0 after stall timeout, closing", stream_id);
break;
}
// Adaptive: cap read to current per-stream target window
let adaptive_cap = remoteingress_protocol::compute_window_for_stream_count(
stream_counter.load(Ordering::Relaxed),
) as usize;
let max_read = w.min(buf.len()).min(adaptive_cap);
tokio::select! {
read_result = up_read.read(&mut buf[..max_read]) => {
@@ -658,6 +681,7 @@ async fn handle_edge_connection(
stream_id,
});
}
stream_counter.fetch_sub(1, Ordering::Relaxed);
});
}
FRAME_DATA => {
@@ -719,8 +743,9 @@ async fn handle_edge_connection(
_ = ping_ticker.tick() => {
let ping_frame = encode_frame(0, FRAME_PING, &[]);
if frame_writer_tx.try_send(ping_frame).is_err() {
log::warn!("Failed to send PING to edge {}, writer channel full/closed", edge_id);
break;
// Control channel full — skip this PING cycle.
// The 45s liveness timeout gives margin for the channel to drain.
log::warn!("PING send to edge {} failed, control channel full — skipping", edge_id);
}
log::trace!("Sent PING to edge {}", edge_id);
}

View File

@@ -32,6 +32,15 @@ pub fn encode_window_update(stream_id: u32, frame_type: u8, increment: u32) -> V
encode_frame(stream_id, frame_type, &increment.to_be_bytes())
}
/// Compute the target per-stream window size based on the number of active streams.
/// Total memory budget is ~32MB shared across all streams. As more streams are active,
/// each gets a smaller window. This adapts to current demand — few streams get high
/// throughput, many streams save memory and reduce control frame pressure.
pub fn compute_window_for_stream_count(active: u32) -> u32 {
let per_stream = (32 * 1024 * 1024u64) / (active.max(1) as u64);
per_stream.clamp(64 * 1024, INITIAL_STREAM_WINDOW as u64) as u32
}
/// Decode a WINDOW_UPDATE payload into a byte increment. Returns None if payload is malformed.
pub fn decode_window_update(payload: &[u8]) -> Option<u32> {
if payload.len() != 4 {

View File

@@ -3,6 +3,6 @@
*/
export const commitinfo = {
name: '@serve.zone/remoteingress',
version: '4.5.9',
version: '4.6.0',
description: 'Edge ingress tunnel for DcRouter - accepts incoming TCP connections at network edge and tunnels them to DcRouter SmartProxy preserving client IP via PROXY protocol v1.'
}