v4.5.4

changelog.md

@@ -1,5 +1,38 @@
 # Changelog
 
+## 2026-03-15 - 4.5.4 - fix(remoteingress-core)
+preserve stream close ordering and add flow-control stall timeouts
+
+- Send CLOSE and CLOSE_BACK frames on the data channel so they arrive after the final stream data frames.
+- Log and abort stalled upload and download paths when flow-control windows stay empty for 120 seconds.
+- Apply a 60-second timeout when writing buffered stream data to the upstream connection to prevent hung streams.
+
+## 2026-03-15 - 4.5.3 - fix(remoteingress-core)
+prioritize control frames over data in edge and hub tunnel writers
+
+- Split tunnel/frame writers into separate control and data channels in edge and hub
+- Use biased select loops so PING, PONG, WINDOW_UPDATE, OPEN, and CLOSE frames are sent before data frames
+- Route stream data through dedicated data channels while keeping OPEN, CLOSE, and flow-control updates on control channels to prevent keepalive starvation under load
+
+## 2026-03-15 - 4.5.2 - fix(remoteingress-core)
+improve stream flow control retries and increase channel buffer capacity
+
+- increase per-stream mpsc channel capacity from 128 to 256 on both edge and hub paths
+- only reset accumulated window update bytes after a successful try_send to avoid dropping flow-control credits when the update channel is busy
+
+## 2026-03-15 - 4.5.1 - fix(protocol)
+increase per-stream flow control window and channel buffers to improve high-RTT throughput
+
+- raise the initial stream window from 256 KB to 4 MB to allow more in-flight data per stream
+- increase edge and hub mpsc channel capacities from 16 to 128 to better absorb throughput under flow control
+
+## 2026-03-15 - 4.5.0 - feat(remoteingress-core)
+add per-stream flow control for edge and hub tunnel data transfer
+
+- introduce WINDOW_UPDATE frame types and protocol helpers for per-stream flow control
+- track per-stream send windows on both edge and hub to limit reads based on available capacity
+- send window updates after downstream writes to reduce channel pressure during large transfers
+
 ## 2026-03-15 - 4.4.1 - fix(remoteingress-core)
 prevent stream data loss by applying backpressure and closing saturated channels
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@serve.zone/remoteingress",
-  "version": "4.4.1",
+  "version": "4.5.4",
   "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",

rust/crates/remoteingress-core/src/edge.rs

@@ -4,7 +4,7 @@ use std::sync::Arc;
 use std::time::Duration;
 use tokio::io::{AsyncBufReadExt, AsyncReadExt, AsyncWriteExt, BufReader};
 use tokio::net::{TcpListener, TcpStream};
-use tokio::sync::{mpsc, Mutex, RwLock};
+use tokio::sync::{mpsc, Mutex, Notify, RwLock};
 use tokio::task::JoinHandle;
 use tokio::time::{Instant, sleep_until};
 use tokio_rustls::TlsConnector;
@@ -13,6 +13,17 @@ use serde::{Deserialize, Serialize};
 
 use remoteingress_protocol::*;
 
+/// Per-stream state tracked in the edge's client_writers map.
+struct EdgeStreamState {
+    /// Channel to deliver FRAME_DATA_BACK payloads to the hub_to_client task.
+    back_tx: mpsc::Sender<Vec<u8>>,
+    /// Send window for FRAME_DATA (upload direction).
+    /// Decremented by the client reader, incremented by FRAME_WINDOW_UPDATE_BACK from hub.
+    send_window: Arc<AtomicU32>,
+    /// Notifier to wake the client reader when the window opens.
+    window_notify: Arc<Notify>,
+}
+
 /// Edge configuration (hub-host + credentials only; ports come from hub).
 #[derive(Debug, Clone, Deserialize, Serialize)]
 #[serde(rename_all = "camelCase")]
@@ -351,17 +362,32 @@ async fn connect_to_hub_and_run(
         }
     });
 
-    // Client socket map: stream_id -> sender for writing data back to client
+    // Client socket map: stream_id -> per-stream state (back channel + flow control)
-    let client_writers: Arc<Mutex<HashMap<u32, mpsc::Sender<Vec<u8>>>>> =
+    let client_writers: Arc<Mutex<HashMap<u32, EdgeStreamState>>> =
         Arc::new(Mutex::new(HashMap::new()));
 
-    // A5: Channel-based tunnel writer replaces Arc<Mutex<WriteHalf>>
+    // QoS dual-channel tunnel writer: control frames (PONG/WINDOW_UPDATE/CLOSE/OPEN)
-    let (tunnel_writer_tx, mut tunnel_writer_rx) = mpsc::channel::<Vec<u8>>(4096);
+    // have priority over data frames (DATA). Prevents PING starvation under load.
+    let (tunnel_ctrl_tx, mut tunnel_ctrl_rx) = mpsc::channel::<Vec<u8>>(64);
+    let (tunnel_data_tx, mut tunnel_data_rx) = mpsc::channel::<Vec<u8>>(4096);
+    // Legacy alias — control channel for PONG, CLOSE, WINDOW_UPDATE, OPEN
+    let tunnel_writer_tx = tunnel_ctrl_tx.clone();
     let tw_token = connection_token.clone();
     let tunnel_writer_handle = tokio::spawn(async move {
         loop {
             tokio::select! {
-                data = tunnel_writer_rx.recv() => {
+                biased; // control frames always take priority over data
+                ctrl = tunnel_ctrl_rx.recv() => {
+                    match ctrl {
+                        Some(frame_data) => {
+                            if write_half.write_all(&frame_data).await.is_err() {
+                                break;
+                            }
+                        }
+                        None => break,
+                    }
+                }
+                data = tunnel_data_rx.recv() => {
                     match data {
                         Some(frame_data) => {
                             if write_half.write_all(&frame_data).await.is_err() {
@@ -382,6 +408,7 @@ async fn connect_to_hub_and_run(
         &handshake.listen_ports,
         &mut port_listeners,
         &tunnel_writer_tx,
+        &tunnel_data_tx,
         &client_writers,
         active_streams,
         next_stream_id,
@@ -407,17 +434,31 @@ async fn connect_to_hub_and_run(
 
                         match frame.frame_type {
                             FRAME_DATA_BACK => {
-                                // Non-blocking send to prevent head-of-line blocking in the main dispatch loop.
+                                // Non-blocking dispatch to per-stream channel.
-                                // If the per-stream channel is full, close the stream rather than silently
+                                // With flow control, the sender should rarely exceed the channel capacity.
-                                // dropping data (which would corrupt the TCP stream).
                                 let mut writers = client_writers.lock().await;
-                                if let Some(tx) = writers.get(&frame.stream_id) {
+                                if let Some(state) = writers.get(&frame.stream_id) {
-                                    if tx.try_send(frame.payload).is_err() {
+                                    if state.back_tx.try_send(frame.payload).is_err() {
-                                        log::warn!("Stream {} back-channel full, closing stream to prevent data corruption", frame.stream_id);
+                                        log::warn!("Stream {} back-channel full, closing stream", frame.stream_id);
                                         writers.remove(&frame.stream_id);
                                     }
                                 }
                             }
+                            FRAME_WINDOW_UPDATE_BACK => {
+                                // Hub consumed data — increase our send window for this stream (upload direction)
+                                if let Some(increment) = decode_window_update(&frame.payload) {
+                                    if increment > 0 {
+                                        let writers = client_writers.lock().await;
+                                        if let Some(state) = writers.get(&frame.stream_id) {
+                                            let prev = state.send_window.fetch_add(increment, Ordering::Release);
+                                            if prev + increment > MAX_WINDOW_SIZE {
+                                                state.send_window.store(MAX_WINDOW_SIZE, Ordering::Release);
+                                            }
+                                            state.window_notify.notify_one();
+                                        }
+                                    }
+                                }
+                            }
                             FRAME_CLOSE_BACK => {
                                 let mut writers = client_writers.lock().await;
                                 writers.remove(&frame.stream_id);
@@ -433,6 +474,7 @@ async fn connect_to_hub_and_run(
                                         &update.listen_ports,
                                         &mut port_listeners,
                                         &tunnel_writer_tx,
+                                        &tunnel_data_tx,
                                         &client_writers,
                                         active_streams,
                                         next_stream_id,
@@ -494,8 +536,9 @@ async fn connect_to_hub_and_run(
 fn apply_port_config(
     new_ports: &[u16],
     port_listeners: &mut HashMap<u16, JoinHandle<()>>,
-    tunnel_writer_tx: &mpsc::Sender<Vec<u8>>,
+    tunnel_ctrl_tx: &mpsc::Sender<Vec<u8>>,
-    client_writers: &Arc<Mutex<HashMap<u32, mpsc::Sender<Vec<u8>>>>>,
+    tunnel_data_tx: &mpsc::Sender<Vec<u8>>,
+    client_writers: &Arc<Mutex<HashMap<u32, EdgeStreamState>>>,
     active_streams: &Arc<AtomicU32>,
     next_stream_id: &Arc<AtomicU32>,
     edge_id: &str,
@@ -514,7 +557,8 @@ fn apply_port_config(
 
     // Add new ports
     for &port in new_set.difference(&old_set) {
-        let tunnel_writer_tx = tunnel_writer_tx.clone();
+        let tunnel_ctrl_tx = tunnel_ctrl_tx.clone();
+        let tunnel_data_tx = tunnel_data_tx.clone();
         let client_writers = client_writers.clone();
         let active_streams = active_streams.clone();
         let next_stream_id = next_stream_id.clone();
@@ -537,7 +581,8 @@ fn apply_port_config(
                         match accept_result {
                             Ok((client_stream, client_addr)) => {
                                 let stream_id = next_stream_id.fetch_add(1, Ordering::Relaxed);
-                                let tunnel_writer_tx = tunnel_writer_tx.clone();
+                                let tunnel_ctrl_tx = tunnel_ctrl_tx.clone();
+                                let tunnel_data_tx = tunnel_data_tx.clone();
                                 let client_writers = client_writers.clone();
                                 let active_streams = active_streams.clone();
                                 let edge_id = edge_id.clone();
@@ -552,7 +597,8 @@ fn apply_port_config(
                                         stream_id,
                                         port,
                                         &edge_id,
-                                        tunnel_writer_tx,
+                                        tunnel_ctrl_tx,
+                                        tunnel_data_tx,
                                         client_writers,
                                         client_token,
                                     )
@@ -582,8 +628,9 @@ async fn handle_client_connection(
     stream_id: u32,
     dest_port: u16,
     edge_id: &str,
-    tunnel_writer_tx: mpsc::Sender<Vec<u8>>,
+    tunnel_ctrl_tx: mpsc::Sender<Vec<u8>>,
-    client_writers: Arc<Mutex<HashMap<u32, mpsc::Sender<Vec<u8>>>>>,
+    tunnel_data_tx: mpsc::Sender<Vec<u8>>,
+    client_writers: Arc<Mutex<HashMap<u32, EdgeStreamState>>>,
     client_token: CancellationToken,
 ) {
     let client_ip = client_addr.ip().to_string();
@@ -592,33 +639,52 @@ async fn handle_client_connection(
     // Determine edge IP (use 0.0.0.0 as placeholder — hub doesn't use it for routing)
     let edge_ip = "0.0.0.0";
 
-    // Send OPEN frame with PROXY v1 header via writer channel
+    // Send OPEN frame with PROXY v1 header via control channel
     let proxy_header = build_proxy_v1_header(&client_ip, edge_ip, client_port, dest_port);
     let open_frame = encode_frame(stream_id, FRAME_OPEN, proxy_header.as_bytes());
-    if tunnel_writer_tx.send(open_frame).await.is_err() {
+    if tunnel_ctrl_tx.send(open_frame).await.is_err() {
         return;
     }
 
-    // Set up channel for data coming back from hub
+    // Set up channel for data coming back from hub (capacity 16 is sufficient with flow control)
     let (back_tx, mut back_rx) = mpsc::channel::<Vec<u8>>(256);
+    let send_window = Arc::new(AtomicU32::new(INITIAL_STREAM_WINDOW));
+    let window_notify = Arc::new(Notify::new());
     {
         let mut writers = client_writers.lock().await;
-        writers.insert(stream_id, back_tx);
+        writers.insert(stream_id, EdgeStreamState {
+            back_tx,
+            send_window: Arc::clone(&send_window),
+            window_notify: Arc::clone(&window_notify),
+        });
     }
 
     let (mut client_read, mut client_write) = client_stream.into_split();
 
-    // Task: hub -> client
+    // Task: hub -> client (download direction)
+    // 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 hub_to_client = tokio::spawn(async move {
+        let mut consumed_since_update: u32 = 0;
         loop {
             tokio::select! {
                 data = back_rx.recv() => {
                     match data {
                         Some(data) => {
+                            let len = data.len() as u32;
                             if client_write.write_all(&data).await.is_err() {
                                 break;
                             }
+                            // Track consumption for flow control
+                            consumed_since_update += len;
+                            if consumed_since_update >= WINDOW_UPDATE_THRESHOLD {
+                                let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE, consumed_since_update);
+                                if wu_tx.try_send(frame).is_ok() {
+                                    consumed_since_update = 0;
+                                }
+                                // If try_send fails, keep accumulating — retry on next threshold
+                            }
                         }
                         None => break,
                     }
@@ -626,23 +692,45 @@ async fn handle_client_connection(
                 _ = hub_to_client_token.cancelled() => break,
             }
         }
+        // Send final window update for any remaining consumed bytes
+        if consumed_since_update > 0 {
+            let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE, consumed_since_update);
+            let _ = wu_tx.try_send(frame);
+        }
         let _ = client_write.shutdown().await;
     });
 
-    // Task: client -> hub (via writer channel)
+    // Task: client -> hub (upload direction) with per-stream flow control
     let mut buf = vec![0u8; 32768];
     loop {
+        // Wait for send window to have capacity (with stall timeout)
+        loop {
+            let w = send_window.load(Ordering::Acquire);
+            if w > 0 { break; }
+            tokio::select! {
+                _ = window_notify.notified() => continue,
+                _ = client_token.cancelled() => break,
+                _ = tokio::time::sleep(Duration::from_secs(120)) => {
+                    log::warn!("Stream {} upload stalled (window empty for 120s)", stream_id);
+                    break;
+                }
+            }
+        }
+        if client_token.is_cancelled() { break; }
+
+        // Limit read size to available window
+        let w = send_window.load(Ordering::Acquire) as usize;
+        let max_read = w.min(buf.len());
+
         tokio::select! {
-            read_result = client_read.read(&mut buf) => {
+            read_result = client_read.read(&mut buf[..max_read]) => {
                 match read_result {
                     Ok(0) => break,
                     Ok(n) => {
+                        send_window.fetch_sub(n as u32, Ordering::Release);
                         let data_frame = encode_frame(stream_id, FRAME_DATA, &buf[..n]);
-                        // Use send().await for backpressure — this is a per-stream task so
+                        if tunnel_data_tx.send(data_frame).await.is_err() {
-                        // blocking only stalls this stream, not others. Prevents data loss
+                            log::warn!("Stream {} data channel closed, closing", stream_id);
-                        // for large transfers (e.g. 352MB Docker layers).
-                        if tunnel_writer_tx.send(data_frame).await.is_err() {
-                            log::warn!("Stream {} tunnel writer closed, closing", stream_id);
                             break;
                         }
                     }
@@ -653,10 +741,10 @@ async fn handle_client_connection(
         }
     }
 
-    // Send CLOSE frame (only if not cancelled)
+    // Send CLOSE frame via DATA channel (must arrive AFTER last DATA for this stream)
     if !client_token.is_cancelled() {
         let close_frame = encode_frame(stream_id, FRAME_CLOSE, &[]);
-        let _ = tunnel_writer_tx.try_send(close_frame);
+        let _ = tunnel_data_tx.try_send(close_frame);
     }
 
     // Cleanup

rust/crates/remoteingress-core/src/hub.rs

@@ -1,9 +1,10 @@
 use std::collections::HashMap;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicU32, Ordering};
 use std::time::Duration;
 use tokio::io::{AsyncBufReadExt, AsyncReadExt, AsyncWriteExt, BufReader};
 use tokio::net::{TcpListener, TcpStream};
-use tokio::sync::{mpsc, Mutex, RwLock, Semaphore};
+use tokio::sync::{mpsc, Mutex, Notify, RwLock, Semaphore};
 use tokio::time::{interval, sleep_until, Instant};
 use tokio_rustls::TlsAcceptor;
 use tokio_util::sync::CancellationToken;
@@ -11,6 +12,19 @@ use serde::{Deserialize, Serialize};
 
 use remoteingress_protocol::*;
 
+/// Per-stream state tracked in the hub's stream map.
+struct HubStreamState {
+    /// Channel to deliver FRAME_DATA payloads to the upstream writer task.
+    data_tx: mpsc::Sender<Vec<u8>>,
+    /// Cancellation token for this stream.
+    cancel_token: CancellationToken,
+    /// Send window for FRAME_DATA_BACK (download direction).
+    /// Decremented by the upstream reader, incremented by FRAME_WINDOW_UPDATE from edge.
+    send_window: Arc<AtomicU32>,
+    /// Notifier to wake the upstream reader when the window opens.
+    window_notify: Arc<Notify>,
+}
+
 /// Hub configuration.
 #[derive(Debug, Clone, Deserialize, Serialize)]
 #[serde(rename_all = "camelCase")]
@@ -109,7 +123,7 @@ pub struct TunnelHub {
 struct ConnectedEdgeInfo {
     connected_at: u64,
     peer_addr: String,
-    active_streams: Arc<Mutex<HashMap<u32, (mpsc::Sender<Vec<u8>>, CancellationToken)>>>,
+    active_streams: Arc<Mutex<HashMap<u32, HubStreamState>>>,
     config_tx: mpsc::Sender<EdgeConfigUpdate>,
     #[allow(dead_code)] // kept alive for Drop — cancels child tokens when edge is removed
     cancel_token: CancellationToken,
@@ -333,7 +347,7 @@ async fn handle_edge_connection(
     write_half.write_all(handshake_json.as_bytes()).await?;
 
     // Track this edge
-    let streams: Arc<Mutex<HashMap<u32, (mpsc::Sender<Vec<u8>>, CancellationToken)>>> =
+    let streams: Arc<Mutex<HashMap<u32, HubStreamState>>> =
         Arc::new(Mutex::new(HashMap::new()));
     let now = std::time::SystemTime::now()
         .duration_since(std::time::UNIX_EPOCH)
@@ -357,14 +371,28 @@ async fn handle_edge_connection(
         );
     }
 
-    // A5: Channel-based writer replaces Arc<Mutex<WriteHalf>>
+    // QoS dual-channel tunnel writer: control frames (PING/PONG/WINDOW_UPDATE/CLOSE)
-    // All frame writes go through this channel → dedicated writer task serializes them
+    // have priority over data frames (DATA_BACK). This prevents PING starvation under load.
-    let (frame_writer_tx, mut frame_writer_rx) = mpsc::channel::<Vec<u8>>(4096);
+    let (ctrl_tx, mut ctrl_rx) = mpsc::channel::<Vec<u8>>(64);
+    let (data_tx, mut data_rx) = mpsc::channel::<Vec<u8>>(4096);
+    // Legacy alias for code that sends both control and data (will be migrated)
+    let frame_writer_tx = ctrl_tx.clone();
     let writer_token = edge_token.clone();
     let writer_handle = tokio::spawn(async move {
         loop {
             tokio::select! {
-                data = frame_writer_rx.recv() => {
+                biased; // control frames always take priority over data
+                ctrl = ctrl_rx.recv() => {
+                    match ctrl {
+                        Some(frame_data) => {
+                            if write_half.write_all(&frame_data).await.is_err() {
+                                break;
+                            }
+                        }
+                        None => break,
+                    }
+                }
+                data = data_rx.recv() => {
                     match data {
                         Some(frame_data) => {
                             if write_half.write_all(&frame_data).await.is_err() {
@@ -453,7 +481,8 @@ async fn handle_edge_connection(
                                 let edge_id_clone = edge_id.clone();
                                 let event_tx_clone = event_tx.clone();
                                 let streams_clone = streams.clone();
-                                let writer_tx = frame_writer_tx.clone();
+                                let writer_tx = ctrl_tx.clone();   // control: CLOSE_BACK, WINDOW_UPDATE_BACK
+                                let data_writer_tx = data_tx.clone(); // data: DATA_BACK
                                 let target = target_host.clone();
                                 let stream_token = edge_token.child_token();
 
@@ -462,11 +491,18 @@ async fn handle_edge_connection(
                                     stream_id,
                                 });
 
-                                // Create channel for data from edge to this stream
+                                // Create channel for data from edge to this stream (capacity 16 is sufficient with flow control)
                                 let (data_tx, mut data_rx) = mpsc::channel::<Vec<u8>>(256);
+                                let send_window = Arc::new(AtomicU32::new(INITIAL_STREAM_WINDOW));
+                                let window_notify = Arc::new(Notify::new());
                                 {
                                     let mut s = streams.lock().await;
-                                    s.insert(stream_id, (data_tx, stream_token.clone()));
+                                    s.insert(stream_id, HubStreamState {
+                                        data_tx,
+                                        cancel_token: stream_token.clone(),
+                                        send_window: Arc::clone(&send_window),
+                                        window_notify: Arc::clone(&window_notify),
+                                    });
                                 }
 
                                 // Spawn task: connect to SmartProxy, send PROXY header, pipe data
@@ -490,15 +526,36 @@ async fn handle_edge_connection(
                                             upstream.into_split();
 
                                         // Forward data from edge (via channel) to SmartProxy
+                                        // After writing to upstream, send WINDOW_UPDATE_BACK to edge
                                         let writer_token = stream_token.clone();
+                                        let wub_tx = writer_tx.clone();
                                         let writer_for_edge_data = tokio::spawn(async move {
+                                            let mut consumed_since_update: u32 = 0;
                                             loop {
                                                 tokio::select! {
                                                     data = data_rx.recv() => {
                                                         match data {
                                                             Some(data) => {
-                                                                if up_write.write_all(&data).await.is_err() {
+                                                                let len = data.len() as u32;
-                                                                    break;
+                                                                match tokio::time::timeout(
+                                                                    Duration::from_secs(60),
+                                                                    up_write.write_all(&data),
+                                                                ).await {
+                                                                    Ok(Ok(())) => {}
+                                                                    Ok(Err(_)) => break,
+                                                                    Err(_) => {
+                                                                        log::warn!("Stream {} write to upstream timed out (60s)", stream_id);
+                                                                        break;
+                                                                    }
+                                                                }
+                                                                // Track consumption for flow control
+                                                                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);
+                                                                    if wub_tx.try_send(frame).is_ok() {
+                                                                        consumed_since_update = 0;
+                                                                    }
+                                                                    // If try_send fails, keep accumulating — retry on next threshold
                                                                 }
                                                             }
                                                             None => break,
@@ -507,24 +564,47 @@ async fn handle_edge_connection(
                                                     _ = writer_token.cancelled() => break,
                                                 }
                                             }
+                                            // Send final window update for remaining consumed bytes
+                                            if consumed_since_update > 0 {
+                                                let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE_BACK, consumed_since_update);
+                                                let _ = wub_tx.try_send(frame);
+                                            }
                                             let _ = up_write.shutdown().await;
                                         });
 
                                         // Forward data from SmartProxy back to edge via writer channel
+                                        // with per-stream flow control (check send_window before reading)
                                         let mut buf = vec![0u8; 32768];
                                         loop {
+                                            // Wait for send window to have capacity (with stall timeout)
+                                            loop {
+                                                let w = send_window.load(Ordering::Acquire);
+                                                if w > 0 { break; }
+                                                tokio::select! {
+                                                    _ = window_notify.notified() => continue,
+                                                    _ = stream_token.cancelled() => break,
+                                                    _ = tokio::time::sleep(Duration::from_secs(120)) => {
+                                                        log::warn!("Stream {} download stalled (window empty for 120s)", stream_id);
+                                                        break;
+                                                    }
+                                                }
+                                            }
+                                            if stream_token.is_cancelled() { break; }
+
+                                            // Limit read size to available window
+                                            let w = send_window.load(Ordering::Acquire) as usize;
+                                            let max_read = w.min(buf.len());
+
                                             tokio::select! {
-                                                read_result = up_read.read(&mut buf) => {
+                                                read_result = up_read.read(&mut buf[..max_read]) => {
                                                     match read_result {
                                                         Ok(0) => break,
                                                         Ok(n) => {
+                                                            send_window.fetch_sub(n as u32, Ordering::Release);
                                                             let frame =
                                                                 encode_frame(stream_id, FRAME_DATA_BACK, &buf[..n]);
-                                                            // Use send().await for backpressure — this is a per-stream task so
+                                                            if data_writer_tx.send(frame).await.is_err() {
-                                                            // blocking only stalls this stream, not others. Prevents data loss
+                                                                log::warn!("Stream {} data channel closed, closing", stream_id);
-                                                            // for large transfers (e.g. 352MB Docker layers).
-                                                            if writer_tx.send(frame).await.is_err() {
-                                                                log::warn!("Stream {} writer channel closed, closing", stream_id);
                                                                 break;
                                                             }
                                                         }
@@ -535,10 +615,10 @@ async fn handle_edge_connection(
                                             }
                                         }
 
-                                        // Send CLOSE_BACK to edge (only if not cancelled)
+                                        // Send CLOSE_BACK via DATA channel (must arrive AFTER last DATA_BACK)
                                         if !stream_token.is_cancelled() {
                                             let close_frame = encode_frame(stream_id, FRAME_CLOSE_BACK, &[]);
-                                            let _ = writer_tx.try_send(close_frame);
+                                            let _ = data_writer_tx.try_send(close_frame);
                                         }
 
                                         writer_for_edge_data.abort();
@@ -548,10 +628,10 @@ async fn handle_edge_connection(
 
                                     if let Err(e) = result {
                                         log::error!("Stream {} error: {}", stream_id, e);
-                                        // Send CLOSE_BACK on error (only if not cancelled)
+                                        // Send CLOSE_BACK via DATA channel on error (must arrive after any DATA_BACK)
                                         if !stream_token.is_cancelled() {
                                             let close_frame = encode_frame(stream_id, FRAME_CLOSE_BACK, &[]);
-                                            let _ = writer_tx.try_send(close_frame);
+                                            let _ = data_writer_tx.try_send(close_frame);
                                         }
                                     }
 
@@ -569,23 +649,37 @@ async fn handle_edge_connection(
                                 });
                             }
                             FRAME_DATA => {
-                                // Non-blocking send to prevent head-of-line blocking in the main dispatch loop.
+                                // Non-blocking dispatch to per-stream channel.
-                                // If the per-stream channel is full, close the stream rather than silently
+                                // With flow control, the sender should rarely exceed the channel capacity.
-                                // dropping data (which would corrupt the TCP stream).
                                 let mut s = streams.lock().await;
-                                if let Some((tx, _)) = s.get(&frame.stream_id) {
+                                if let Some(state) = s.get(&frame.stream_id) {
-                                    if tx.try_send(frame.payload).is_err() {
+                                    if state.data_tx.try_send(frame.payload).is_err() {
-                                        log::warn!("Stream {} data channel full, closing stream to prevent data corruption", frame.stream_id);
+                                        log::warn!("Stream {} data channel full, closing stream", frame.stream_id);
-                                        if let Some((_, token)) = s.remove(&frame.stream_id) {
+                                        if let Some(state) = s.remove(&frame.stream_id) {
-                                            token.cancel();
+                                            state.cancel_token.cancel();
+                                        }
+                                    }
+                                }
+                            }
+                            FRAME_WINDOW_UPDATE => {
+                                // Edge consumed data — increase our send window for this stream
+                                if let Some(increment) = decode_window_update(&frame.payload) {
+                                    if increment > 0 {
+                                        let s = streams.lock().await;
+                                        if let Some(state) = s.get(&frame.stream_id) {
+                                            let prev = state.send_window.fetch_add(increment, Ordering::Release);
+                                            if prev + increment > MAX_WINDOW_SIZE {
+                                                state.send_window.store(MAX_WINDOW_SIZE, Ordering::Release);
+                                            }
+                                            state.window_notify.notify_one();
                                         }
                                     }
                                 }
                             }
                             FRAME_CLOSE => {
                                 let mut s = streams.lock().await;
-                                if let Some((_, token)) = s.remove(&frame.stream_id) {
+                                if let Some(state) = s.remove(&frame.stream_id) {
-                                    token.cancel();
+                                    state.cancel_token.cancel();
                                     let _ = event_tx.try_send(HubEvent::StreamClosed {
                                         edge_id: edge_id.clone(),
                                         stream_id: frame.stream_id,

rust/crates/remoteingress-protocol/src/lib.rs

@@ -9,6 +9,8 @@ pub const FRAME_CLOSE_BACK: u8 = 0x05;
 pub const FRAME_CONFIG: u8 = 0x06;     // Hub -> Edge: configuration update
 pub const FRAME_PING: u8 = 0x07;       // Hub -> Edge: heartbeat probe
 pub const FRAME_PONG: u8 = 0x08;       // Edge -> Hub: heartbeat response
+pub const FRAME_WINDOW_UPDATE: u8 = 0x09;      // Edge -> Hub: per-stream flow control
+pub const FRAME_WINDOW_UPDATE_BACK: u8 = 0x0A;  // Hub -> Edge: per-stream flow control
 
 // Frame header size: 4 (stream_id) + 1 (type) + 4 (length) = 9 bytes
 pub const FRAME_HEADER_SIZE: usize = 9;
@@ -16,6 +18,28 @@ pub const FRAME_HEADER_SIZE: usize = 9;
 // Maximum payload size (16 MB)
 pub const MAX_PAYLOAD_SIZE: u32 = 16 * 1024 * 1024;
 
+// Per-stream flow control constants
+/// Initial per-stream window size (4 MB). Sized for full throughput at high RTT:
+/// at 100ms RTT, this sustains ~40 MB/s per stream.
+pub const INITIAL_STREAM_WINDOW: u32 = 4 * 1024 * 1024;
+/// Send WINDOW_UPDATE after consuming this many bytes (half the initial window).
+pub const WINDOW_UPDATE_THRESHOLD: u32 = INITIAL_STREAM_WINDOW / 2;
+/// Maximum window size to prevent overflow.
+pub const MAX_WINDOW_SIZE: u32 = 16 * 1024 * 1024;
+
+/// Encode a WINDOW_UPDATE frame for a specific stream.
+pub fn encode_window_update(stream_id: u32, frame_type: u8, increment: u32) -> Vec<u8> {
+    encode_frame(stream_id, frame_type, &increment.to_be_bytes())
+}
+
+/// 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 {
+        return None;
+    }
+    Some(u32::from_be_bytes([payload[0], payload[1], payload[2], payload[3]]))
+}
+
 /// A single multiplexed frame.
 #[derive(Debug, Clone)]
 pub struct Frame {

ts/00_commitinfo_data.ts

@@ -3,6 +3,6 @@
  */
 export const commitinfo = {
   name: '@serve.zone/remoteingress',
-  version: '4.4.1',
+  version: '4.5.4',
   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.'
 }