v25.11.6

changelog.md

@@ -1,5 +1,77 @@
 # Changelog
 
+## 2026-03-16 - 25.11.6 - fix(rustproxy-http,rustproxy-passthrough)
+improve upstream connection cleanup and graceful tunnel shutdown
+
+- Evict pooled HTTP/2 connections when their driver exits and shorten the maximum pooled H2 age to reduce reuse of stale upstream connections.
+- Strip hop-by-hop headers from backend responses before forwarding to HTTP/2 clients to avoid invalid H2 response handling.
+- Replace immediate task aborts in WebSocket and TCP tunnel watchdogs with cancellation-driven graceful shutdown plus timed fallback aborts.
+- Use non-blocking semaphore acquisition in the TCP listener so connection limits do not stall the accept loop for the entire port.
+
+## 2026-03-16 - 25.11.5 - fix(repo)
+no changes to commit
+
+
+## 2026-03-15 - 25.11.4 - fix(rustproxy-http)
+report streamed HTTP and WebSocket bytes per chunk for real-time throughput metrics
+
+- Update CountingBody to record bytes immediately on each data frame instead of aggregating until completion or drop
+- Record WebSocket tunnel traffic inside both copy loops and remove the final aggregate byte report to keep throughput metrics current
+
+## 2026-03-15 - 25.11.3 - fix(repo)
+no changes to commit
+
+
+## 2026-03-15 - 25.11.2 - fix(rustproxy-http)
+avoid reusing HTTP/1 senders during streaming responses and relax HTTP/2 keep-alive timeouts
+
+- Stop returning HTTP/1 senders to the connection pool before upstream response bodies finish streaming to prevent unsafe reuse on active connections.
+- Increase HTTP/2 keep-alive timeout from 5 seconds to 30 seconds in proxy connection builders to better support longer-lived backend streams.
+- Improves reliability for large streaming payloads and backend fallback request handling.
+
+## 2026-03-15 - 25.11.1 - fix(rustproxy-http)
+keep connection idle tracking alive during streaming and tune HTTP/2 connection lifetimes
+
+- Propagate connection activity tracking through HTTP/1, HTTP/2, and WebSocket forwarding so active request and response body streams do not trigger the idle watchdog.
+- Update CountingBody to refresh connection activity timestamps while data frames are polled during uploads and downloads.
+- Increase pooled HTTP/2 max age and set explicit HTTP/2 connection window sizes to improve long-lived streaming behavior.
+
+## 2026-03-15 - 25.11.0 - feat(rustproxy-http)
+add HTTP/2 Extended CONNECT WebSocket proxy support
+
+- Enable HTTP/2 CONNECT protocol support on the Hyper auto connection builder
+- Detect WebSocket requests for both HTTP/1 Upgrade and HTTP/2 Extended CONNECT flows
+- Translate HTTP/2 WebSocket requests to an HTTP/1.1 backend handshake and return RFC-compliant client responses
+
+## 2026-03-12 - 25.10.7 - fix(rustproxy-http)
+remove Host header from HTTP/2 upstream requests while preserving it for HTTP/1 retries
+
+- strips the Host header before sending HTTP/2 upstream requests so :authority from the URI is used instead
+- avoids 400 responses from nginx caused by sending both Host and :authority headers
+- keeps a cloned header set for bodyless request retries so HTTP/1 fallback still retains the Host header
+
+## 2026-03-12 - 25.10.6 - fix(rustproxy-http)
+use the requested domain as HTTP/2 authority instead of the backend host and port
+
+- build HTTP/2 absolute URIs from the client-facing domain so the :authority pseudo-header matches the Host header
+- remove backend port from generated HTTP/2 request URIs and fall back to the upstream host only when no domain is available
+- apply the authority handling consistently across pooled, inline, and generic upstream request paths
+
+## 2026-03-12 - 25.10.5 - fix(rustproxy-http)
+configure HTTP/2 client builders with a Tokio timer for keep-alive handling
+
+- Adds TokioTimer to all HTTP/2 client builder instances in proxy_service.
+- Ensures configured HTTP/2 keep-alive interval and timeout settings have the required timer runtime support.
+
+## 2026-03-12 - 25.10.4 - fix(rustproxy-http)
+stabilize upstream HTTP/2 forwarding and fallback behavior
+
+- Remove hop-by-hop headers before forwarding requests to HTTP/2 backends to comply with RFC 9113.
+- Use ALPN-enabled TLS configuration whenever HTTP/2 is possible, including explicit H2 connections and retries.
+- Add HTTP/2 handshake timeouts, tuned connection settings, and fallback to HTTP/1 when H2 negotiation times out or fails.
+- Register pooled HTTP/2 senders only after a successful first request to avoid reusing broken connections.
+- Build absolute URIs for HTTP/2 upstream requests so pseudo-headers such as scheme and authority are derived correctly.
+
 ## 2026-03-12 - 25.10.3 - fix(rustproxy-http)
 include request domain in backend proxy error and protocol detection logs
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@push.rocks/smartproxy",
-  "version": "25.10.3",
+  "version": "25.11.6",
   "private": false,
   "description": "A powerful proxy package with unified route-based configuration for high traffic management. Features include SSL/TLS support, flexible routing patterns, WebSocket handling, advanced security options, and automatic ACME certificate management.",
   "main": "dist_ts/index.js",

rust/crates/rustproxy-http/src/connection_pool.rs

@@ -20,6 +20,7 @@ const IDLE_TIMEOUT: Duration = Duration::from_secs(90);
 const EVICTION_INTERVAL: Duration = Duration::from_secs(30);
 /// Maximum age for pooled HTTP/2 connections before proactive eviction.
 /// Prevents staleness from backends that close idle connections (e.g. nginx GOAWAY).
+/// 120s is well within typical server GOAWAY windows (nginx: ~60s idle, envoy: ~60s).
 const MAX_H2_AGE: Duration = Duration::from_secs(120);
 
 /// Identifies a unique backend endpoint.

rust/crates/rustproxy-http/src/counting_body.rs

@@ -11,20 +11,22 @@ use rustproxy_metrics::MetricsCollector;
 
 /// Wraps any `http_body::Body` and counts data bytes passing through.
 ///
-/// When the body is fully consumed or dropped, accumulated byte counts
-/// are reported to the `MetricsCollector`.
+/// Each chunk is reported to the `MetricsCollector` immediately so that
+/// the throughput tracker (sampled at 1 Hz) reflects real-time data flow.
 ///
 /// The inner body is pinned on the heap to support `!Unpin` types like `hyper::body::Incoming`.
 pub struct CountingBody<B> {
     inner: Pin<Box<B>>,
-    counted_bytes: AtomicU64,
     metrics: Arc<MetricsCollector>,
     route_id: Option<String>,
     source_ip: Option<String>,
     /// Whether we count bytes as "in" (request body) or "out" (response body).
     direction: Direction,
-    /// Whether we've already reported the bytes (to avoid double-reporting on drop).
-    reported: bool,
+    /// Optional connection-level activity tracker. When set, poll_frame updates this
+    /// to keep the idle watchdog alive during active body streaming (uploads/downloads).
+    connection_activity: Option<Arc<AtomicU64>>,
+    /// Start instant for computing elapsed ms for connection_activity.
+    activity_start: Option<std::time::Instant>,
 }
 
 /// Which direction the bytes flow.
@@ -47,42 +49,36 @@ impl<B> CountingBody<B> {
     ) -> Self {
         Self {
             inner: Box::pin(inner),
-            counted_bytes: AtomicU64::new(0),
             metrics,
             route_id,
             source_ip,
             direction,
-            reported: false,
+            connection_activity: None,
+            activity_start: None,
         }
     }
 
-    /// Report accumulated bytes to the metrics collector.
-    fn report(&mut self) {
-        if self.reported {
-            return;
-        }
-        self.reported = true;
-
-        let bytes = self.counted_bytes.load(Ordering::Relaxed);
-        if bytes == 0 {
-            return;
-        }
+    /// Set the connection-level activity tracker. When set, each data frame
+    /// updates this timestamp to prevent the idle watchdog from killing the
+    /// connection during active body streaming.
+    pub fn with_connection_activity(mut self, activity: Arc<AtomicU64>, start: std::time::Instant) -> Self {
+        self.connection_activity = Some(activity);
+        self.activity_start = Some(start);
+        self
+    }
 
+    /// Report a chunk of bytes immediately to the metrics collector.
+    #[inline]
+    fn report_chunk(&self, len: u64) {
         let route_id = self.route_id.as_deref();
         let source_ip = self.source_ip.as_deref();
         match self.direction {
-            Direction::In => self.metrics.record_bytes(bytes, 0, route_id, source_ip),
-            Direction::Out => self.metrics.record_bytes(0, bytes, route_id, source_ip),
+            Direction::In => self.metrics.record_bytes(len, 0, route_id, source_ip),
+            Direction::Out => self.metrics.record_bytes(0, len, route_id, source_ip),
         }
     }
 }
 
-impl<B> Drop for CountingBody<B> {
-    fn drop(&mut self) {
-        self.report();
-    }
-}
-
 // CountingBody is Unpin because inner is Pin<Box<B>> (always Unpin).
 impl<B> Unpin for CountingBody<B> {}
 
@@ -102,16 +98,18 @@ where
         match this.inner.as_mut().poll_frame(cx) {
             Poll::Ready(Some(Ok(frame))) => {
                 if let Some(data) = frame.data_ref() {
-                    this.counted_bytes.fetch_add(data.len() as u64, Ordering::Relaxed);
+                    let len = data.len() as u64;
+                    // Report bytes immediately so the 1 Hz throughput sampler sees them
+                    this.report_chunk(len);
+                    // Keep the connection-level idle watchdog alive during body streaming
+                    if let (Some(activity), Some(start)) = (&this.connection_activity, &this.activity_start) {
+                        activity.store(start.elapsed().as_millis() as u64, Ordering::Relaxed);
+                    }
                 }
                 Poll::Ready(Some(Ok(frame)))
             }
             Poll::Ready(Some(Err(e))) => Poll::Ready(Some(Err(e))),
-            Poll::Ready(None) => {
-                // Body is fully consumed — report now
-                this.report();
-                Poll::Ready(None)
-            }
+            Poll::Ready(None) => Poll::Ready(None),
             Poll::Pending => Poll::Pending,
         }
     }

rust/crates/rustproxy-http/src/proxy_service.rs

@@ -33,6 +33,14 @@ use crate::request_filter::RequestFilter;
 use crate::response_filter::ResponseFilter;
 use crate::upstream_selector::UpstreamSelector;
 
+/// Per-connection context for keeping the idle watchdog alive during body streaming.
+/// Passed through the forwarding chain so CountingBody can update the timestamp.
+#[derive(Clone)]
+struct ConnActivity {
+    last_activity: Arc<AtomicU64>,
+    start: std::time::Instant,
+}
+
 /// Default upstream connect timeout (30 seconds).
 const DEFAULT_CONNECT_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(30);
 
@@ -294,8 +302,9 @@ impl HttpProxyService {
             let cn = cancel_inner.clone();
             let la = Arc::clone(&la_inner);
             let st = start;
+            let ca = ConnActivity { last_activity: Arc::clone(&la_inner), start };
             async move {
-                let result = svc.handle_request(req, peer, port, cn).await;
+                let result = svc.handle_request(req, peer, port, cn, ca).await;
                 // Mark request end — update activity timestamp before guard drops
                 la.store(st.elapsed().as_millis() as u64, Ordering::Relaxed);
                 drop(req_guard); // Explicitly drop to decrement active_requests
@@ -304,8 +313,13 @@ impl HttpProxyService {
         });
 
         // Auto-detect h1 vs h2 based on ALPN / connection preface.
-        // serve_connection_with_upgrades supports h1 Upgrade (WebSocket) and h2 CONNECT.
-        let builder = hyper_util::server::conn::auto::Builder::new(hyper_util::rt::TokioExecutor::new());
+        // serve_connection_with_upgrades supports h1 Upgrade (WebSocket) and h2 Extended CONNECT (RFC 8441).
+        let mut builder = hyper_util::server::conn::auto::Builder::new(hyper_util::rt::TokioExecutor::new());
+        // Configure H2 server settings: Extended CONNECT for WebSocket + flow control tuning
+        builder.http2()
+            .enable_connect_protocol()
+            .initial_stream_window_size(2 * 1024 * 1024)        // 2MB per stream (vs default 64KB)
+            .initial_connection_window_size(8 * 1024 * 1024);   // 8MB per client connection
         let conn = builder.serve_connection_with_upgrades(io, service);
         // Pin on the heap — auto::UpgradeableConnection is !Unpin
         let mut conn = Box::pin(conn);
@@ -365,6 +379,7 @@ impl HttpProxyService {
         peer_addr: std::net::SocketAddr,
         port: u16,
         cancel: CancellationToken,
+        conn_activity: ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         let host = req.headers()
             .get("host")
@@ -482,16 +497,23 @@ impl HttpProxyService {
         let domain_str = host.as_deref().unwrap_or("-");
         self.upstream_selector.connection_started(&upstream_key);
 
-        // Check for WebSocket upgrade
-        let is_websocket = req.headers()
+        // Check for WebSocket upgrade: H1 (Upgrade header) or H2 Extended CONNECT (RFC 8441)
+        let is_h1_websocket = req.headers()
             .get("upgrade")
             .and_then(|v| v.to_str().ok())
             .map(|v| v.eq_ignore_ascii_case("websocket"))
             .unwrap_or(false);
 
-        if is_websocket {
+        let is_h2_websocket = req.method() == hyper::Method::CONNECT
+            && req.extensions()
+                .get::<hyper::ext::Protocol>()
+                .map(|p| p.as_str().eq_ignore_ascii_case("websocket"))
+                .unwrap_or(false);
+
+        if is_h1_websocket || is_h2_websocket {
             let result = self.handle_websocket_upgrade(
-                req, peer_addr, &upstream, route_match.route, route_id, &upstream_key, cancel, &ip_str,
+                req, peer_addr, &upstream, route_match.route, route_id, &upstream_key, cancel, &ip_str, is_h2_websocket,
+                if is_h2_websocket { Some(conn_activity.clone()) } else { None },
             ).await;
             // Note: for WebSocket, connection_ended is called inside
             // the spawned tunnel task when the connection closes.
@@ -539,6 +561,15 @@ impl HttpProxyService {
             }
         }
 
+        // Remove hop-by-hop headers (RFC 9113 §8.2.2 forbids connection-specific headers in H2)
+        upstream_headers.remove("connection");
+        upstream_headers.remove("keep-alive");
+        upstream_headers.remove("proxy-connection");
+        upstream_headers.remove("transfer-encoding");
+        upstream_headers.remove("te");
+        upstream_headers.remove("trailer");
+        upstream_headers.remove("upgrade");
+
         // Add standard reverse-proxy headers (X-Forwarded-*)
         {
             let original_host = host.as_deref().unwrap_or("");
@@ -623,7 +654,7 @@ impl HttpProxyService {
                     self.metrics.set_backend_protocol(&upstream_key, "h2");
                     let result = self.forward_h2_pooled(
                         sender, parts, body, upstream_headers, &upstream_path,
-                        route_match.route, route_id, &ip_str, &pool_key, domain_str,
+                        route_match.route, route_id, &ip_str, &pool_key, domain_str, &conn_activity,
                     ).await;
                     self.upstream_selector.connection_ended(&upstream_key);
                     return result;
@@ -634,8 +665,8 @@ impl HttpProxyService {
         // --- Fresh connection path ---
         self.metrics.backend_pool_miss(&upstream_key);
 
-        // Choose TLS config: use ALPN config for auto-detect probe, plain config otherwise
-        let tls_config = if needs_alpn_probe {
+        // Choose TLS config: use ALPN config when H2 is possible (auto-detect probe OR explicit H2)
+        let tls_config = if needs_alpn_probe || use_h2 {
             &self.backend_tls_config_alpn
         } else {
             &self.backend_tls_config
@@ -762,19 +793,19 @@ impl HttpProxyService {
                 self.forward_h2_with_fallback(
                     io, parts, body, upstream_headers, &upstream_path,
                     &upstream, route_match.route, route_id, &ip_str, &final_pool_key,
-                    host.clone(), domain_str,
+                    host.clone(), domain_str, &conn_activity,
                 ).await
             } else {
                 // Explicit H2 mode: hard-fail on handshake error (preserved behavior)
                 self.forward_h2(
                     io, parts, body, upstream_headers, &upstream_path,
-                    &upstream, route_match.route, route_id, &ip_str, &final_pool_key, domain_str,
+                    &upstream, route_match.route, route_id, &ip_str, &final_pool_key, domain_str, &conn_activity,
                 ).await
             }
         } else {
             self.forward_h1(
                 io, parts, body, upstream_headers, &upstream_path,
-                &upstream, route_match.route, route_id, &ip_str, &final_pool_key, domain_str,
+                &upstream, route_match.route, route_id, &ip_str, &final_pool_key, domain_str, &conn_activity,
             ).await
         };
         self.upstream_selector.connection_ended(&upstream_key);
@@ -797,6 +828,7 @@ impl HttpProxyService {
         source_ip: &str,
         pool_key: &crate::connection_pool::PoolKey,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         let backend_key = format!("{}:{}", pool_key.host, pool_key.port);
 
@@ -805,7 +837,7 @@ impl HttpProxyService {
             self.metrics.backend_pool_hit(&backend_key);
             return self.forward_h1_with_sender(
                 pooled_sender, parts, body, upstream_headers, upstream_path,
-                route, route_id, source_ip, pool_key, domain,
+                route, route_id, source_ip, pool_key, domain, conn_activity,
             ).await;
         }
 
@@ -828,7 +860,7 @@ impl HttpProxyService {
             }
         });
 
-        self.forward_h1_with_sender(sender, parts, body, upstream_headers, upstream_path, route, route_id, source_ip, pool_key, domain).await
+        self.forward_h1_with_sender(sender, parts, body, upstream_headers, upstream_path, route, route_id, source_ip, pool_key, domain, conn_activity).await
     }
 
     /// Common H1 forwarding logic used by both fresh and pooled paths.
@@ -844,6 +876,7 @@ impl HttpProxyService {
         source_ip: &str,
         pool_key: &crate::connection_pool::PoolKey,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         // Always use HTTP/1.1 for h1 backend connections (h2 incoming requests have version HTTP/2.0)
         let mut upstream_req = Request::builder()
@@ -862,7 +895,7 @@ impl HttpProxyService {
             route_id.map(|s| s.to_string()),
             Some(source_ip.to_string()),
             Direction::In,
-        );
+        ).with_connection_activity(Arc::clone(&conn_activity.last_activity), conn_activity.start);
         let boxed_body: BoxBody<Bytes, hyper::Error> = BoxBody::new(counting_req_body);
 
         let upstream_req = upstream_req.body(boxed_body).unwrap();
@@ -877,10 +910,17 @@ impl HttpProxyService {
             }
         };
 
-        // Return sender to pool (body streams lazily, sender is reusable once response head is received)
-        self.connection_pool.checkin_h1(pool_key.clone(), sender);
+        // Note: we do NOT return the sender to the pool here because the response body
+        // hasn't been fully streamed yet. Pooling a sender while its response body is still
+        // in-flight risks another request being dispatched on the same connection if is_ready()
+        // momentarily returns true between chunks. The sender is dropped after this scope,
+        // and the backend connection remains alive via the spawned conn driver task until
+        // the response body finishes streaming.
+        // For small/empty responses, the sender could theoretically be reused, but the safety
+        // of large streaming responses (e.g. 352MB Docker layers) takes priority.
+        drop(sender);
 
-        self.build_streaming_response(upstream_response, route, route_id, source_ip).await
+        self.build_streaming_response(upstream_response, route, route_id, source_ip, conn_activity).await
     }
 
     /// Forward request to backend via HTTP/2 with body streaming (fresh connection).
@@ -898,32 +938,54 @@ impl HttpProxyService {
         source_ip: &str,
         pool_key: &crate::connection_pool::PoolKey,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         let backend_key = format!("{}:{}", pool_key.host, pool_key.port);
         let exec = hyper_util::rt::TokioExecutor::new();
-        // Explicitly type the handshake with BoxBody for uniform pool type
+        let mut h2_builder = hyper::client::conn::http2::Builder::new(exec);
+        h2_builder
+            .timer(hyper_util::rt::TokioTimer::new())
+            .keep_alive_interval(std::time::Duration::from_secs(10))
+            .keep_alive_timeout(std::time::Duration::from_secs(30))
+            .initial_stream_window_size(2 * 1024 * 1024)
+            .initial_connection_window_size(16 * 1024 * 1024);
         let (sender, conn): (
             hyper::client::conn::http2::SendRequest<BoxBody<Bytes, hyper::Error>>,
             hyper::client::conn::http2::Connection<TokioIo<BackendStream>, BoxBody<Bytes, hyper::Error>, hyper_util::rt::TokioExecutor>,
-        ) = match hyper::client::conn::http2::handshake(exec, io).await {
-            Ok(h) => h,
-            Err(e) => {
+        ) = match tokio::time::timeout(self.connect_timeout, h2_builder.handshake(io)).await {
+            Ok(Ok(h)) => h,
+            Ok(Err(e)) => {
                 error!(backend = %backend_key, domain = %domain, error = %e, "Backend H2 handshake failed");
                 self.metrics.backend_handshake_error(&backend_key);
                 return Ok(error_response(StatusCode::BAD_GATEWAY, "Backend H2 handshake failed"));
             }
+            Err(_) => {
+                error!(backend = %backend_key, domain = %domain, "Backend H2 handshake timeout");
+                self.metrics.backend_handshake_error(&backend_key);
+                return Ok(error_response(StatusCode::GATEWAY_TIMEOUT, "Backend H2 handshake timeout"));
+            }
         };
 
-        tokio::spawn(async move {
-            if let Err(e) = conn.await {
-                debug!("HTTP/2 upstream connection error: {}", e);
-            }
-        });
+        // Spawn the H2 connection driver; proactively evict from pool on exit
+        // so the next request gets a fresh connection instead of a dead sender.
+        {
+            let pool = Arc::clone(&self.connection_pool);
+            let key = pool_key.clone();
+            tokio::spawn(async move {
+                if let Err(e) = conn.await {
+                    debug!("HTTP/2 upstream connection error: {}", e);
+                }
+                pool.remove_h2(&key);
+            });
+        }
 
-        // Register for multiplexed reuse
-        self.connection_pool.register_h2(pool_key.clone(), sender.clone());
-
-        self.forward_h2_with_sender(sender, parts, body, upstream_headers, upstream_path, route, route_id, source_ip, Some(pool_key), domain).await
+        // Clone sender for potential pool registration; register only after first request succeeds
+        let sender_for_pool = sender.clone();
+        let result = self.forward_h2_with_sender(sender, parts, body, upstream_headers, upstream_path, route, route_id, source_ip, Some(pool_key), domain, conn_activity).await;
+        if matches!(&result, Ok(ref resp) if resp.status() != StatusCode::BAD_GATEWAY) {
+            self.connection_pool.register_h2(pool_key.clone(), sender_for_pool);
+        }
+        result
     }
 
     /// Forward request using an existing (pooled) HTTP/2 sender.
@@ -941,6 +1003,7 @@ impl HttpProxyService {
         source_ip: &str,
         pool_key: &crate::connection_pool::PoolKey,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         // Save retry state for bodyless requests (cheap: Method is an enum, HeaderMap clones Arc-backed Bytes)
         let retry_state = if body.is_end_stream() {
@@ -951,7 +1014,7 @@ impl HttpProxyService {
 
         let result = self.forward_h2_with_sender(
             sender, parts, body, upstream_headers, upstream_path,
-            route, route_id, source_ip, Some(pool_key), domain,
+            route, route_id, source_ip, Some(pool_key), domain, conn_activity,
         ).await;
 
         // If the request failed (502) and we can retry with an empty body, do so
@@ -962,7 +1025,7 @@ impl HttpProxyService {
                     "Stale pooled H2 sender, retrying with fresh connection");
                 return self.retry_h2_with_fresh_connection(
                     method, headers, upstream_path,
-                    pool_key, route, route_id, source_ip, domain,
+                    pool_key, route, route_id, source_ip, domain, conn_activity,
                 ).await;
             }
         }
@@ -981,6 +1044,7 @@ impl HttpProxyService {
         route_id: Option<&str>,
         source_ip: &str,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         let backend_key = format!("{}:{}", pool_key.host, pool_key.port);
 
@@ -989,7 +1053,7 @@ impl HttpProxyService {
         let backend = if pool_key.use_tls {
             match tokio::time::timeout(
                 self.connect_timeout,
-                connect_tls_backend(&self.backend_tls_config, &pool_key.host, pool_key.port),
+                connect_tls_backend(&self.backend_tls_config_alpn, &pool_key.host, pool_key.port),
             ).await {
                 Ok(Ok(tls)) => BackendStream::Tls(tls),
                 Ok(Err(e)) => {
@@ -1028,31 +1092,55 @@ impl HttpProxyService {
 
         let io = TokioIo::new(backend);
         let exec = hyper_util::rt::TokioExecutor::new();
+        let mut h2_builder = hyper::client::conn::http2::Builder::new(exec);
+        h2_builder
+            .timer(hyper_util::rt::TokioTimer::new())
+            .keep_alive_interval(std::time::Duration::from_secs(10))
+            .keep_alive_timeout(std::time::Duration::from_secs(30))
+            .initial_stream_window_size(2 * 1024 * 1024)
+            .initial_connection_window_size(16 * 1024 * 1024);
         let (mut sender, conn): (
             hyper::client::conn::http2::SendRequest<BoxBody<Bytes, hyper::Error>>,
             hyper::client::conn::http2::Connection<TokioIo<BackendStream>, BoxBody<Bytes, hyper::Error>, hyper_util::rt::TokioExecutor>,
-        ) = match hyper::client::conn::http2::handshake(exec, io).await {
-            Ok(h) => h,
-            Err(e) => {
+        ) = match tokio::time::timeout(self.connect_timeout, h2_builder.handshake(io)).await {
+            Ok(Ok(h)) => h,
+            Ok(Err(e)) => {
                 error!(backend = %backend_key, domain = %domain, error = %e, "H2 retry: handshake failed");
                 self.metrics.backend_handshake_error(&backend_key);
+                self.metrics.backend_connection_closed(&backend_key);
                 return Ok(error_response(StatusCode::BAD_GATEWAY, "Backend H2 retry handshake failed"));
             }
+            Err(_) => {
+                error!(backend = %backend_key, domain = %domain, "H2 retry: handshake timeout");
+                self.metrics.backend_handshake_error(&backend_key);
+                self.metrics.backend_connection_closed(&backend_key);
+                return Ok(error_response(StatusCode::GATEWAY_TIMEOUT, "Backend H2 retry handshake timeout"));
+            }
         };
 
-        tokio::spawn(async move {
-            if let Err(e) = conn.await {
-                debug!("H2 retry: upstream connection error: {}", e);
-            }
-        });
+        // Spawn the H2 connection driver; proactively evict from pool on exit.
+        {
+            let pool = Arc::clone(&self.connection_pool);
+            let key = pool_key.clone();
+            tokio::spawn(async move {
+                if let Err(e) = conn.await {
+                    debug!("H2 retry: upstream connection error: {}", e);
+                }
+                pool.remove_h2(&key);
+            });
+        }
 
-        // Register fresh sender in pool for future requests
-        self.connection_pool.register_h2(pool_key.clone(), sender.clone());
-
-        // Build request with empty body
+        // Build request with empty body using absolute URI for H2 pseudo-headers
+        let scheme = if pool_key.use_tls { "https" } else { "http" };
+        let authority = if domain != "-" { domain } else { pool_key.host.as_str() };
+        let h2_uri = format!("{}://{}{}", scheme, authority, upstream_path);
         let mut upstream_req = Request::builder()
             .method(method)
-            .uri(upstream_path);
+            .uri(&h2_uri);
+
+        // Remove Host header for H2 — :authority pseudo-header (from URI) is sufficient
+        let mut upstream_headers = upstream_headers;
+        upstream_headers.remove("host");
 
         if let Some(headers) = upstream_req.headers_mut() {
             *headers = upstream_headers;
@@ -1065,7 +1153,9 @@ impl HttpProxyService {
 
         match sender.send_request(upstream_req).await {
             Ok(resp) => {
-                let result = self.build_streaming_response(resp, route, route_id, source_ip).await;
+                // Register in pool only after request succeeds
+                self.connection_pool.register_h2(pool_key.clone(), sender);
+                let result = self.build_streaming_response(resp, route, route_id, source_ip, conn_activity).await;
                 // Close the fresh backend connection (opened above)
                 self.metrics.backend_connection_closed(&backend_key);
                 result
@@ -1073,7 +1163,6 @@ impl HttpProxyService {
             Err(e) => {
                 error!(backend = %backend_key, domain = %domain, error = %e, "H2 retry: request failed");
                 self.metrics.backend_request_error(&backend_key);
-                self.connection_pool.remove_h2(pool_key);
                 // Close the fresh backend connection (opened above)
                 self.metrics.backend_connection_closed(&backend_key);
                 Ok(error_response(StatusCode::BAD_GATEWAY, "Backend H2 request failed on retry"))
@@ -1093,7 +1182,7 @@ impl HttpProxyService {
         io: TokioIo<BackendStream>,
         parts: hyper::http::request::Parts,
         body: Incoming,
-        upstream_headers: hyper::HeaderMap,
+        mut upstream_headers: hyper::HeaderMap,
         upstream_path: &str,
         upstream: &crate::upstream_selector::UpstreamSelection,
         route: &rustproxy_config::RouteConfig,
@@ -1102,34 +1191,94 @@ impl HttpProxyService {
         pool_key: &crate::connection_pool::PoolKey,
         requested_host: Option<String>,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         let exec = hyper_util::rt::TokioExecutor::new();
-        let handshake_result: Result<(
-            hyper::client::conn::http2::SendRequest<BoxBody<Bytes, hyper::Error>>,
-            hyper::client::conn::http2::Connection<TokioIo<BackendStream>, BoxBody<Bytes, hyper::Error>, hyper_util::rt::TokioExecutor>,
-        ), hyper::Error> = hyper::client::conn::http2::handshake(exec, io).await;
+        let mut h2_builder = hyper::client::conn::http2::Builder::new(exec);
+        h2_builder
+            .timer(hyper_util::rt::TokioTimer::new())
+            .keep_alive_interval(std::time::Duration::from_secs(10))
+            .keep_alive_timeout(std::time::Duration::from_secs(30))
+            .initial_stream_window_size(2 * 1024 * 1024)
+            .initial_connection_window_size(16 * 1024 * 1024);
+        let handshake_result = tokio::time::timeout(
+            self.connect_timeout,
+            h2_builder.handshake(io),
+        ).await;
 
         match handshake_result {
-            Ok((mut sender, conn)) => {
-                tokio::spawn(async move {
-                    if let Err(e) = conn.await {
-                        debug!("HTTP/2 upstream connection error: {}", e);
+            Err(_) => {
+                // H2 handshake timed out — fall back to H1
+                let bk = format!("{}:{}", upstream.host, upstream.port);
+                warn!(
+                    backend = %bk,
+                    domain = %domain,
+                    "H2 handshake timeout, falling back to H1"
+                );
+                self.metrics.backend_h2_failure(&bk);
+                self.metrics.backend_handshake_error(&bk);
+
+                let cache_key = crate::protocol_cache::ProtocolCacheKey {
+                    host: upstream.host.clone(),
+                    port: upstream.port,
+                    requested_host: requested_host.clone(),
+                };
+                self.protocol_cache.insert(cache_key, crate::protocol_cache::DetectedProtocol::H1);
+
+                match self.reconnect_backend(upstream, domain).await {
+                    Some(fallback_backend) => {
+                        let h1_pool_key = crate::connection_pool::PoolKey {
+                            host: upstream.host.clone(),
+                            port: upstream.port,
+                            use_tls: upstream.use_tls,
+                            h2: false,
+                        };
+                        let fallback_io = TokioIo::new(fallback_backend);
+                        let result = self.forward_h1(
+                            fallback_io, parts, body, upstream_headers, upstream_path,
+                            upstream, route, route_id, source_ip, &h1_pool_key, domain, conn_activity,
+                        ).await;
+                        self.metrics.backend_connection_closed(&bk);
+                        result
                     }
-                });
+                    None => {
+                        Ok(error_response(StatusCode::BAD_GATEWAY, "Backend unavailable after H2 timeout fallback"))
+                    }
+                }
+            }
+            Ok(Ok((mut sender, conn))) => {
+                // Spawn the H2 connection driver; proactively evict from pool on exit.
+                {
+                    let pool = Arc::clone(&self.connection_pool);
+                    let key = pool_key.clone();
+                    tokio::spawn(async move {
+                        if let Err(e) = conn.await {
+                            debug!("HTTP/2 upstream connection error: {}", e);
+                        }
+                        pool.remove_h2(&key);
+                    });
+                }
 
                 // Save retry state before consuming parts/body (for bodyless requests like GET)
+                // Clone BEFORE removing Host — H1 fallback needs Host header
                 let retry_state = if body.is_end_stream() {
                     Some((parts.method.clone(), upstream_headers.clone()))
                 } else {
                     None
                 };
 
+                // Remove Host header for H2 — :authority pseudo-header (from URI) is sufficient
+                upstream_headers.remove("host");
+
                 // Build and send the h2 request inline (don't register in pool yet —
                 // we need to verify the request actually succeeds first, because some
                 // backends advertise h2 via ALPN but don't speak the h2 binary protocol).
+                let scheme = if upstream.use_tls { "https" } else { "http" };
+                let authority = if domain != "-" { domain } else { upstream.host.as_str() };
+                let h2_uri = format!("{}://{}{}", scheme, authority, upstream_path);
                 let mut upstream_req = Request::builder()
                     .method(parts.method)
-                    .uri(upstream_path);
+                    .uri(&h2_uri);
 
                 if let Some(headers) = upstream_req.headers_mut() {
                     *headers = upstream_headers;
@@ -1141,7 +1290,7 @@ impl HttpProxyService {
                     route_id.map(|s| s.to_string()),
                     Some(source_ip.to_string()),
                     Direction::In,
-                );
+                ).with_connection_activity(Arc::clone(&conn_activity.last_activity), conn_activity.start);
                 let boxed_body: BoxBody<Bytes, hyper::Error> = BoxBody::new(counting_req_body);
                 let upstream_req = upstream_req.body(boxed_body).unwrap();
 
@@ -1149,7 +1298,7 @@ impl HttpProxyService {
                     Ok(upstream_response) => {
                         // H2 works! Register sender in pool for multiplexed reuse
                         self.connection_pool.register_h2(pool_key.clone(), sender);
-                        self.build_streaming_response(upstream_response, route, route_id, source_ip).await
+                        self.build_streaming_response(upstream_response, route, route_id, source_ip, conn_activity).await
                     }
                     Err(e) => {
                         // H2 request failed — backend advertises h2 via ALPN but doesn't
@@ -1182,7 +1331,7 @@ impl HttpProxyService {
                                     let fallback_io = TokioIo::new(fallback_backend);
                                     let result = self.forward_h1_empty_body(
                                         fallback_io, method, headers, upstream_path,
-                                        route, route_id, source_ip, &h1_pool_key, domain,
+                                        route, route_id, source_ip, &h1_pool_key, domain, conn_activity,
                                     ).await;
                                     // Close the reconnected backend connection (opened in reconnect_backend)
                                     self.metrics.backend_connection_closed(&bk);
@@ -1198,7 +1347,7 @@ impl HttpProxyService {
                     }
                 }
             }
-            Err(e) => {
+            Ok(Err(e)) => {
                 // H2 handshake truly failed — fall back to H1
                 // Body is NOT consumed yet, so we can retry the full request.
                 let bk = format!("{}:{}", upstream.host, upstream.port);
@@ -1231,7 +1380,7 @@ impl HttpProxyService {
                         let fallback_io = TokioIo::new(fallback_backend);
                         let result = self.forward_h1(
                             fallback_io, parts, body, upstream_headers, upstream_path,
-                            upstream, route, route_id, source_ip, &h1_pool_key, domain,
+                            upstream, route, route_id, source_ip, &h1_pool_key, domain, conn_activity,
                         ).await;
                         // Close the reconnected backend connection (opened in reconnect_backend)
                         self.metrics.backend_connection_closed(&bk);
@@ -1258,6 +1407,7 @@ impl HttpProxyService {
         source_ip: &str,
         pool_key: &crate::connection_pool::PoolKey,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         let backend_key = format!("{}:{}", pool_key.host, pool_key.port);
         let (mut sender, conn): (
@@ -1301,10 +1451,10 @@ impl HttpProxyService {
             }
         };
 
-        // Return sender to pool for keep-alive reuse
-        self.connection_pool.checkin_h1(pool_key.clone(), sender);
+        // Don't pool the sender while response body is still streaming (same safety as forward_h1_with_sender)
+        drop(sender);
 
-        self.build_streaming_response(upstream_response, route, route_id, source_ip).await
+        self.build_streaming_response(upstream_response, route, route_id, source_ip, conn_activity).await
     }
 
     /// Reconnect to a backend (used for H2→H1 fallback).
@@ -1375,10 +1525,23 @@ impl HttpProxyService {
         source_ip: &str,
         pool_key: Option<&crate::connection_pool::PoolKey>,
         domain: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
+        // Build absolute URI for H2 pseudo-headers (:scheme, :authority)
+        // Use the requested domain as authority (not backend address) so :authority matches Host header
+        let scheme = if pool_key.map(|pk| pk.use_tls).unwrap_or(false) { "https" } else { "http" };
+        let authority = if domain != "-" { domain } else {
+            pool_key.map(|pk| pk.host.as_str()).unwrap_or("localhost")
+        };
+        let h2_uri = format!("{}://{}{}", scheme, authority, upstream_path);
         let mut upstream_req = Request::builder()
             .method(parts.method)
-            .uri(upstream_path);
+            .uri(&h2_uri);
+
+        // Remove Host header for H2 — :authority pseudo-header (from URI) is sufficient
+        // Having both Host and :authority causes nginx to return 400
+        let mut upstream_headers = upstream_headers;
+        upstream_headers.remove("host");
 
         if let Some(headers) = upstream_req.headers_mut() {
             *headers = upstream_headers;
@@ -1391,7 +1554,7 @@ impl HttpProxyService {
             route_id.map(|s| s.to_string()),
             Some(source_ip.to_string()),
             Direction::In,
-        );
+        ).with_connection_activity(Arc::clone(&conn_activity.last_activity), conn_activity.start);
         let boxed_body: BoxBody<Bytes, hyper::Error> = BoxBody::new(counting_req_body);
 
         let upstream_req = upstream_req.body(boxed_body).unwrap();
@@ -1412,7 +1575,7 @@ impl HttpProxyService {
             }
         };
 
-        self.build_streaming_response(upstream_response, route, route_id, source_ip).await
+        self.build_streaming_response(upstream_response, route, route_id, source_ip, conn_activity).await
     }
 
     /// Build the client-facing response from an upstream response, streaming the body.
@@ -1425,6 +1588,7 @@ impl HttpProxyService {
         route: &rustproxy_config::RouteConfig,
         route_id: Option<&str>,
         source_ip: &str,
+        conn_activity: &ConnActivity,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         let (resp_parts, resp_body) = upstream_response.into_parts();
 
@@ -1433,6 +1597,19 @@ impl HttpProxyService {
 
         if let Some(headers) = response.headers_mut() {
             *headers = resp_parts.headers;
+
+            // Strip hop-by-hop headers from the backend response.
+            // RFC 9113 §8.2.2 forbids connection-specific headers in HTTP/2 responses;
+            // forwarding them from an H1 backend can cause H2 stream resets.
+            // Mirrors the request-path stripping at the forward methods above.
+            headers.remove("connection");
+            headers.remove("keep-alive");
+            headers.remove("proxy-connection");
+            headers.remove("transfer-encoding");
+            headers.remove("te");
+            headers.remove("trailer");
+            // Note: "upgrade" is intentionally kept — needed for WebSocket 101 responses.
+
             ResponseFilter::apply_headers(route, headers, None);
         }
 
@@ -1445,14 +1622,14 @@ impl HttpProxyService {
             route_id.map(|s| s.to_string()),
             Some(source_ip.to_string()),
             Direction::Out,
-        );
+        ).with_connection_activity(Arc::clone(&conn_activity.last_activity), conn_activity.start);
 
         let body: BoxBody<Bytes, hyper::Error> = BoxBody::new(counting_body);
 
         Ok(response.body(body).unwrap())
     }
 
-    /// Handle a WebSocket upgrade request.
+    /// Handle a WebSocket upgrade request (H1 Upgrade or H2 Extended CONNECT per RFC 8441).
     async fn handle_websocket_upgrade(
         &self,
         req: Request<Incoming>,
@@ -1463,6 +1640,8 @@ impl HttpProxyService {
         upstream_key: &str,
         cancel: CancellationToken,
         source_ip: &str,
+        is_h2: bool,
+        conn_activity: Option<ConnActivity>,
     ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
         use tokio::io::{AsyncReadExt, AsyncWriteExt};
 
@@ -1548,9 +1727,11 @@ impl HttpProxyService {
 
         let (parts, _body) = req.into_parts();
 
+        // H2 Extended CONNECT uses method=CONNECT, but the H1.1 backend expects GET
+        let backend_method = if is_h2 { "GET" } else { parts.method.as_str() };
         let mut raw_request = format!(
             "{} {} HTTP/1.1\r\n",
-            parts.method, upstream_path
+            backend_method, upstream_path
         );
 
         // Copy all original headers (preserving the client's Host header).
@@ -1578,6 +1759,23 @@ impl HttpProxyService {
             }
         }
 
+        // H2 Extended CONNECT doesn't carry H1 WebSocket handshake headers;
+        // inject them so the H1.1 backend can complete the upgrade.
+        if is_h2 {
+            if !parts.headers.contains_key("upgrade") {
+                raw_request.push_str("upgrade: websocket\r\n");
+            }
+            if !parts.headers.contains_key("connection") {
+                raw_request.push_str("connection: Upgrade\r\n");
+            }
+            if !parts.headers.contains_key("sec-websocket-version") {
+                raw_request.push_str("sec-websocket-version: 13\r\n");
+            }
+            if !parts.headers.contains_key("sec-websocket-key") {
+                raw_request.push_str("sec-websocket-key: dGhlIHNhbXBsZSBub25jZQ==\r\n");
+            }
+        }
+
         // Add standard reverse-proxy headers (X-Forwarded-*)
         {
             let original_host = parts.headers.get("host")
@@ -1680,8 +1878,12 @@ impl HttpProxyService {
             ));
         }
 
-        let mut client_resp = Response::builder()
-            .status(StatusCode::SWITCHING_PROTOCOLS);
+        // H1: 101 Switching Protocols; H2: 200 OK (RFC 8441 — hyper requires 2xx for Extended CONNECT upgrade)
+        let mut client_resp = if is_h2 {
+            Response::builder().status(StatusCode::OK)
+        } else {
+            Response::builder().status(StatusCode::SWITCHING_PROTOCOLS)
+        };
 
         if let Some(resp_headers) = client_resp.headers_mut() {
             for line in response_str.lines().skip(1) {
@@ -1692,6 +1894,17 @@ impl HttpProxyService {
                 if let Some((name, value)) = line.split_once(':') {
                     let name = name.trim();
                     let value = value.trim();
+                    // Skip hop-by-hop headers for H2 (forbidden by RFC 9113 §8.2.2)
+                    if is_h2 {
+                        let name_lower = name.to_lowercase();
+                        if name_lower == "upgrade" || name_lower == "connection"
+                            || name_lower == "sec-websocket-accept"
+                            || name_lower == "transfer-encoding"
+                            || name_lower == "keep-alive"
+                        {
+                            continue;
+                        }
+                    }
                     if let Ok(header_name) = hyper::header::HeaderName::from_bytes(name.as_bytes()) {
                         if let Ok(header_value) = hyper::header::HeaderValue::from_str(value) {
                             resp_headers.insert(header_name, header_value);
@@ -1732,48 +1945,89 @@ impl HttpProxyService {
             let last_activity = Arc::new(AtomicU64::new(0));
             let start = std::time::Instant::now();
 
+            // Per-connection cancellation token: the watchdog cancels this instead of
+            // aborting tasks, so the copy loops can shut down gracefully (TLS close_notify).
+            let ws_cancel = CancellationToken::new();
+
+            // For H2 WebSocket: also update the connection-level activity tracker
+            // to prevent the idle watchdog from killing the H2 connection
+            let conn_act_c2u = conn_activity.as_ref().map(|ca| (Arc::clone(&ca.last_activity), ca.start));
+            let conn_act_u2c = conn_activity.as_ref().map(|ca| (Arc::clone(&ca.last_activity), ca.start));
+
             let la1 = Arc::clone(&last_activity);
+            let metrics_c2u = Arc::clone(&metrics);
+            let route_c2u = route_id_owned.clone();
+            let ip_c2u = source_ip_owned.clone();
+            let wsc1 = ws_cancel.clone();
             let c2u = tokio::spawn(async move {
                 let mut buf = vec![0u8; 65536];
                 let mut total = 0u64;
                 loop {
-                    let n = match cr.read(&mut buf).await {
-                        Ok(0) | Err(_) => break,
-                        Ok(n) => n,
+                    let n = tokio::select! {
+                        result = cr.read(&mut buf) => match result {
+                            Ok(0) | Err(_) => break,
+                            Ok(n) => n,
+                        },
+                        _ = wsc1.cancelled() => break,
                     };
                     if uw.write_all(&buf[..n]).await.is_err() {
                         break;
                     }
                     total += n as u64;
+                    metrics_c2u.record_bytes(n as u64, 0, route_c2u.as_deref(), Some(&ip_c2u));
                     la1.store(start.elapsed().as_millis() as u64, Ordering::Relaxed);
+                    if let Some((ref ca, ca_start)) = conn_act_c2u {
+                        ca.store(ca_start.elapsed().as_millis() as u64, Ordering::Relaxed);
+                    }
                 }
-                let _ = uw.shutdown().await;
+                // Graceful shutdown with timeout (sends TLS close_notify / TCP FIN)
+                let _ = tokio::time::timeout(
+                    std::time::Duration::from_secs(2),
+                    uw.shutdown(),
+                ).await;
                 total
             });
 
             let la2 = Arc::clone(&last_activity);
+            let metrics_u2c = Arc::clone(&metrics);
+            let route_u2c = route_id_owned.clone();
+            let ip_u2c = source_ip_owned.clone();
+            let wsc2 = ws_cancel.clone();
             let u2c = tokio::spawn(async move {
                 let mut buf = vec![0u8; 65536];
                 let mut total = 0u64;
                 loop {
-                    let n = match ur.read(&mut buf).await {
-                        Ok(0) | Err(_) => break,
-                        Ok(n) => n,
+                    let n = tokio::select! {
+                        result = ur.read(&mut buf) => match result {
+                            Ok(0) | Err(_) => break,
+                            Ok(n) => n,
+                        },
+                        _ = wsc2.cancelled() => break,
                     };
                     if cw.write_all(&buf[..n]).await.is_err() {
                         break;
                     }
                     total += n as u64;
+                    metrics_u2c.record_bytes(0, n as u64, route_u2c.as_deref(), Some(&ip_u2c));
                     la2.store(start.elapsed().as_millis() as u64, Ordering::Relaxed);
+                    if let Some((ref ca, ca_start)) = conn_act_u2c {
+                        ca.store(ca_start.elapsed().as_millis() as u64, Ordering::Relaxed);
+                    }
                 }
-                let _ = cw.shutdown().await;
+                // Graceful shutdown with timeout (sends TLS close_notify / TCP FIN)
+                let _ = tokio::time::timeout(
+                    std::time::Duration::from_secs(2),
+                    cw.shutdown(),
+                ).await;
                 total
             });
 
-            // Watchdog: monitors inactivity, max lifetime, and cancellation
+            // Watchdog: monitors inactivity, max lifetime, and cancellation.
+            // First cancels the per-connection token for graceful shutdown (close_notify/FIN),
+            // then falls back to abort if the tasks are stuck (e.g. on a blocked write_all).
             let la_watch = Arc::clone(&last_activity);
-            let c2u_handle = c2u.abort_handle();
-            let u2c_handle = u2c.abort_handle();
+            let c2u_abort = c2u.abort_handle();
+            let u2c_abort = u2c.abort_handle();
             let inactivity_timeout = ws_inactivity_timeout;
             let max_lifetime = ws_max_lifetime;
 
@@ -1785,8 +2039,6 @@ impl HttpProxyService {
                         _ = tokio::time::sleep(check_interval) => {}
                         _ = cancel.cancelled() => {
                             debug!("WebSocket tunnel cancelled by shutdown");
-                            c2u_handle.abort();
-                            u2c_handle.abort();
                             break;
                         }
                     }
@@ -1794,8 +2046,6 @@ impl HttpProxyService {
                     // Check max lifetime
                     if start.elapsed() >= max_lifetime {
                         debug!("WebSocket tunnel exceeded max lifetime, closing");
-                        c2u_handle.abort();
-                        u2c_handle.abort();
                         break;
                     }
 
@@ -1805,13 +2055,18 @@ impl HttpProxyService {
                         let elapsed_since_activity = start.elapsed().as_millis() as u64 - current;
                         if elapsed_since_activity >= inactivity_timeout.as_millis() as u64 {
                             debug!("WebSocket tunnel inactive for {}ms, closing", elapsed_since_activity);
-                            c2u_handle.abort();
-                            u2c_handle.abort();
                             break;
                         }
                     }
                     last_seen = current;
                 }
+                // Phase 1: Signal copy loops to exit gracefully (allows close_notify/FIN)
+                ws_cancel.cancel();
+                // Phase 2: Wait for graceful shutdown (2s shutdown timeout + 2s margin)
+                tokio::time::sleep(std::time::Duration::from_secs(4)).await;
+                // Phase 3: Force-abort if still stuck (e.g. blocked on write_all)
+                c2u_abort.abort();
+                u2c_abort.abort();
             });
 
             let bytes_in = c2u.await.unwrap_or(0);
@@ -1821,9 +2076,7 @@ impl HttpProxyService {
             debug!("WebSocket tunnel closed: {} bytes in, {} bytes out", bytes_in, bytes_out);
 
             upstream_selector.connection_ended(&upstream_key_owned);
-            if let Some(ref rid) = route_id_owned {
-                metrics.record_bytes(bytes_in, bytes_out, Some(rid.as_str()), Some(&source_ip_owned));
-            }
+            // Bytes already reported per-chunk in the copy loops above
         });
 
         let body: BoxBody<Bytes, hyper::Error> = BoxBody::new(

rust/crates/rustproxy-passthrough/src/forwarder.rs

@@ -97,16 +97,25 @@ pub async fn forward_bidirectional_with_timeouts(
     let last_activity = Arc::new(AtomicU64::new(0));
     let start = std::time::Instant::now();
 
+    // Per-connection cancellation token: the watchdog cancels this instead of
+    // aborting tasks, so the copy loops can shut down gracefully (TCP FIN instead
+    // of RST, TLS close_notify if the stream is TLS-wrapped).
+    let conn_cancel = CancellationToken::new();
+
     let la1 = Arc::clone(&last_activity);
     let initial_len = initial_data.map_or(0u64, |d| d.len() as u64);
     let metrics_c2b = metrics.clone();
+    let cc1 = conn_cancel.clone();
     let c2b = tokio::spawn(async move {
         let mut buf = vec![0u8; 65536];
         let mut total = initial_len;
         loop {
-            let n = match client_read.read(&mut buf).await {
-                Ok(0) | Err(_) => break,
-                Ok(n) => n,
+            let n = tokio::select! {
+                result = client_read.read(&mut buf) => match result {
+                    Ok(0) | Err(_) => break,
+                    Ok(n) => n,
+                },
+                _ = cc1.cancelled() => break,
             };
             if backend_write.write_all(&buf[..n]).await.is_err() {
                 break;
@@ -117,19 +126,27 @@ pub async fn forward_bidirectional_with_timeouts(
                 ctx.collector.record_bytes(n as u64, 0, ctx.route_id.as_deref(), ctx.source_ip.as_deref());
             }
         }
-        let _ = backend_write.shutdown().await;
+        // Graceful shutdown with timeout (sends TCP FIN / TLS close_notify)
+        let _ = tokio::time::timeout(
+            std::time::Duration::from_secs(2),
+            backend_write.shutdown(),
+        ).await;
         total
     });
 
     let la2 = Arc::clone(&last_activity);
     let metrics_b2c = metrics;
+    let cc2 = conn_cancel.clone();
     let b2c = tokio::spawn(async move {
         let mut buf = vec![0u8; 65536];
         let mut total = 0u64;
         loop {
-            let n = match backend_read.read(&mut buf).await {
-                Ok(0) | Err(_) => break,
-                Ok(n) => n,
+            let n = tokio::select! {
+                result = backend_read.read(&mut buf) => match result {
+                    Ok(0) | Err(_) => break,
+                    Ok(n) => n,
+                },
+                _ = cc2.cancelled() => break,
             };
             if client_write.write_all(&buf[..n]).await.is_err() {
                 break;
@@ -140,14 +157,20 @@ pub async fn forward_bidirectional_with_timeouts(
                 ctx.collector.record_bytes(0, n as u64, ctx.route_id.as_deref(), ctx.source_ip.as_deref());
             }
         }
-        let _ = client_write.shutdown().await;
+        // Graceful shutdown with timeout (sends TCP FIN / TLS close_notify)
+        let _ = tokio::time::timeout(
+            std::time::Duration::from_secs(2),
+            client_write.shutdown(),
+        ).await;
         total
     });
 
-    // Watchdog: inactivity, max lifetime, and cancellation
+    // Watchdog: inactivity, max lifetime, and cancellation.
+    // First cancels the per-connection token for graceful shutdown (FIN/close_notify),
+    // then falls back to abort if the tasks are stuck (e.g. on a blocked write_all).
     let la_watch = Arc::clone(&last_activity);
-    let c2b_handle = c2b.abort_handle();
-    let b2c_handle = b2c.abort_handle();
+    let c2b_abort = c2b.abort_handle();
+    let b2c_abort = b2c.abort_handle();
     let watchdog = tokio::spawn(async move {
         let check_interval = std::time::Duration::from_secs(5);
         let mut last_seen = 0u64;
@@ -155,16 +178,12 @@ pub async fn forward_bidirectional_with_timeouts(
             tokio::select! {
                 _ = cancel.cancelled() => {
                     debug!("Connection cancelled by shutdown");
-                    c2b_handle.abort();
-                    b2c_handle.abort();
                     break;
                 }
                 _ = tokio::time::sleep(check_interval) => {
                     // Check max lifetime
                     if start.elapsed() >= max_lifetime {
                         debug!("Connection exceeded max lifetime, closing");
-                        c2b_handle.abort();
-                        b2c_handle.abort();
                         break;
                     }
 
@@ -174,8 +193,6 @@ pub async fn forward_bidirectional_with_timeouts(
                         let elapsed_since_activity = start.elapsed().as_millis() as u64 - current;
                         if elapsed_since_activity >= inactivity_timeout.as_millis() as u64 {
                             debug!("Connection inactive for {}ms, closing", elapsed_since_activity);
-                            c2b_handle.abort();
-                            b2c_handle.abort();
                             break;
                         }
                     }
@@ -183,6 +200,13 @@ pub async fn forward_bidirectional_with_timeouts(
                 }
             }
         }
+        // Phase 1: Signal copy loops to exit gracefully (allows FIN/close_notify)
+        conn_cancel.cancel();
+        // Phase 2: Wait for graceful shutdown (2s shutdown timeout + 2s margin)
+        tokio::time::sleep(std::time::Duration::from_secs(4)).await;
+        // Phase 3: Force-abort if still stuck (e.g. blocked on write_all)
+        c2b_abort.abort();
+        b2c_abort.abort();
     });
 
     let bytes_in = c2b.await.unwrap_or(0);

rust/crates/rustproxy-passthrough/src/tcp_listener.rs

@@ -465,21 +465,19 @@ impl TcpListenerManager {
                         Ok((stream, peer_addr)) => {
                             let ip = peer_addr.ip();
 
-                            // Global connection limit — acquire semaphore permit with timeout
-                            let permit = match tokio::time::timeout(
-                                std::time::Duration::from_secs(5),
-                                conn_semaphore.clone().acquire_owned(),
-                            ).await {
-                                Ok(Ok(permit)) => permit,
-                                Ok(Err(_)) => {
-                                    // Semaphore closed — shouldn't happen, but be safe
-                                    debug!("Connection semaphore closed, dropping connection from {}", peer_addr);
+                            // Global connection limit — non-blocking check.
+                            // MUST NOT block the accept loop: a blocking acquire would stall
+                            // ALL connections to this port (not just the one over limit), because
+                            // listener.accept() is not polled while we await the semaphore.
+                            let permit = match conn_semaphore.clone().try_acquire_owned() {
+                                Ok(permit) => permit,
+                                Err(tokio::sync::TryAcquireError::NoPermits) => {
+                                    debug!("Global connection limit reached, dropping connection from {}", peer_addr);
                                     drop(stream);
                                     continue;
                                 }
-                                Err(_) => {
-                                    // Timeout — global limit reached
-                                    debug!("Global connection limit reached, dropping connection from {}", peer_addr);
+                                Err(tokio::sync::TryAcquireError::Closed) => {
+                                    debug!("Connection semaphore closed, dropping connection from {}", peer_addr);
                                     drop(stream);
                                     continue;
                                 }
@@ -1396,15 +1394,24 @@ impl TcpListenerManager {
         let last_activity = Arc::new(AtomicU64::new(0));
         let start = std::time::Instant::now();
 
+        // Per-connection cancellation token: the watchdog cancels this instead of
+        // aborting tasks, so the copy loops can shut down gracefully (TLS close_notify
+        // for terminate/reencrypt mode, TCP FIN for passthrough mode).
+        let conn_cancel = CancellationToken::new();
+
         let la1 = Arc::clone(&last_activity);
         let metrics_c2b = metrics.clone();
+        let cc1 = conn_cancel.clone();
         let c2b = tokio::spawn(async move {
             let mut buf = vec![0u8; 65536];
             let mut total = 0u64;
             loop {
-                let n = match client_read.read(&mut buf).await {
-                    Ok(0) | Err(_) => break,
-                    Ok(n) => n,
+                let n = tokio::select! {
+                    result = client_read.read(&mut buf) => match result {
+                        Ok(0) | Err(_) => break,
+                        Ok(n) => n,
+                    },
+                    _ = cc1.cancelled() => break,
                 };
                 if backend_write.write_all(&buf[..n]).await.is_err() {
                     break;
@@ -1418,19 +1425,27 @@ impl TcpListenerManager {
                     ctx.collector.record_bytes(n as u64, 0, ctx.route_id.as_deref(), ctx.source_ip.as_deref());
                 }
             }
-            let _ = backend_write.shutdown().await;
+            // Graceful shutdown with timeout (sends TLS close_notify / TCP FIN)
+            let _ = tokio::time::timeout(
+                std::time::Duration::from_secs(2),
+                backend_write.shutdown(),
+            ).await;
             total
         });
 
         let la2 = Arc::clone(&last_activity);
         let metrics_b2c = metrics;
+        let cc2 = conn_cancel.clone();
         let b2c = tokio::spawn(async move {
             let mut buf = vec![0u8; 65536];
             let mut total = 0u64;
             loop {
-                let n = match backend_read.read(&mut buf).await {
-                    Ok(0) | Err(_) => break,
-                    Ok(n) => n,
+                let n = tokio::select! {
+                    result = backend_read.read(&mut buf) => match result {
+                        Ok(0) | Err(_) => break,
+                        Ok(n) => n,
+                    },
+                    _ = cc2.cancelled() => break,
                 };
                 if client_write.write_all(&buf[..n]).await.is_err() {
                     break;
@@ -1444,14 +1459,20 @@ impl TcpListenerManager {
                     ctx.collector.record_bytes(0, n as u64, ctx.route_id.as_deref(), ctx.source_ip.as_deref());
                 }
             }
-            let _ = client_write.shutdown().await;
+            // Graceful shutdown with timeout (sends TLS close_notify / TCP FIN)
+            let _ = tokio::time::timeout(
+                std::time::Duration::from_secs(2),
+                client_write.shutdown(),
+            ).await;
             total
         });
 
-        // Watchdog task: check for inactivity, max lifetime, and cancellation
+        // Watchdog task: check for inactivity, max lifetime, and cancellation.
+        // First cancels the per-connection token for graceful shutdown (close_notify/FIN),
+        // then falls back to abort if the tasks are stuck (e.g. on a blocked write_all).
         let la_watch = Arc::clone(&last_activity);
-        let c2b_handle = c2b.abort_handle();
-        let b2c_handle = b2c.abort_handle();
+        let c2b_abort = c2b.abort_handle();
+        let b2c_abort = b2c.abort_handle();
         let watchdog = tokio::spawn(async move {
             let check_interval = std::time::Duration::from_secs(5);
             let mut last_seen = 0u64;
@@ -1459,16 +1480,12 @@ impl TcpListenerManager {
                 tokio::select! {
                     _ = cancel.cancelled() => {
                         debug!("Split-stream connection cancelled by shutdown");
-                        c2b_handle.abort();
-                        b2c_handle.abort();
                         break;
                     }
                     _ = tokio::time::sleep(check_interval) => {
                         // Check max lifetime
                         if start.elapsed() >= max_lifetime {
                             debug!("Connection exceeded max lifetime, closing");
-                            c2b_handle.abort();
-                            b2c_handle.abort();
                             break;
                         }
 
@@ -1479,8 +1496,6 @@ impl TcpListenerManager {
                             let elapsed_since_activity = start.elapsed().as_millis() as u64 - current;
                             if elapsed_since_activity >= inactivity_timeout.as_millis() as u64 {
                                 debug!("Connection inactive for {}ms, closing", elapsed_since_activity);
-                                c2b_handle.abort();
-                                b2c_handle.abort();
                                 break;
                             }
                         }
@@ -1488,6 +1503,13 @@ impl TcpListenerManager {
                     }
                 }
             }
+            // Phase 1: Signal copy loops to exit gracefully (allows close_notify/FIN)
+            conn_cancel.cancel();
+            // Phase 2: Wait for graceful shutdown (2s shutdown timeout + 2s margin)
+            tokio::time::sleep(std::time::Duration::from_secs(4)).await;
+            // Phase 3: Force-abort if still stuck (e.g. blocked on write_all)
+            c2b_abort.abort();
+            b2c_abort.abort();
         });
 
         let bytes_in = c2b.await.unwrap_or(0);

ts/00_commitinfo_data.ts

@@ -3,6 +3,6 @@
  */
 export const commitinfo = {
   name: '@push.rocks/smartproxy',
-  version: '25.10.3',
+  version: '25.11.6',
   description: 'A powerful proxy package with unified route-based configuration for high traffic management. Features include SSL/TLS support, flexible routing patterns, WebSocket handling, advanced security options, and automatic ACME certificate management.'
 }