v25.11.14

- Add HTTP keep-alive idle timeout (60s default) with periodic watchdog that
  skips active requests (panic-safe via RAII ActiveRequestGuard)
- Make WebSocket inactivity/max-lifetime timeouts configurable from ConnectionConfig
  instead of hardcoded 1h/24h
- Replace bare copy_bidirectional in socket handler relay with timeout+cancel-aware
  split forwarding (inactivity, max lifetime, graceful shutdown)
- Add CancellationToken to forward_bidirectional_split_with_timeouts so TLS-terminated
  TCP connections respond to graceful shutdown
- Fix graceful_stop to actually abort listener tasks that exceed the shutdown deadline
  (previously they detached and ran forever)
- Add 10s metadata parsing timeout on TS socket-handler-server to prevent stuck sockets

changelog.md

@@ -1,5 +1,208 @@
 # Changelog
 
+## 2026-03-16 - 25.11.14 - fix(rustproxy-http)
+forward vectored write support in ShutdownOnDrop AsyncWrite wrapper
+
+- Implements poll_write_vectored by delegating to the wrapped writer
+- Exposes is_write_vectored so the wrapper preserves underlying AsyncWrite capabilities
+
+## 2026-03-16 - 25.11.13 - fix(rustproxy-http)
+remove hot-path debug logging from HTTP/1 connection pool hits
+
+- Stops emitting debug logs when reusing HTTP/1 idle connections in the connection pool.
+- Keeps pool hit behavior unchanged while reducing overhead on a frequently executed path.
+
+## 2026-03-16 - 25.11.12 - fix(rustproxy-http)
+remove connection pool hit logging and keep logging limited to actual failures
+
+- Removes debug and warning logs for HTTP/2 connection pool hits and age checks.
+- Keeps pool behavior unchanged while reducing noisy per-request logging in the Rust HTTP proxy layer.
+
+## 2026-03-16 - 25.11.11 - fix(rustproxy-http)
+improve HTTP/2 proxy error logging with warning-level connection failures and debug error details
+
+- Adds debug-formatted error fields to HTTP/2 handshake, retry, fallback, and request failure logs
+- Promotes upstream HTTP/2 connection error logs from debug to warn to improve operational visibility
+
+## 2026-03-16 - 25.11.10 - fix(rustproxy-http)
+validate pooled HTTP/2 connections asynchronously before reuse and evict stale senders
+
+- Add an async ready() check with a 500ms timeout before reusing pooled HTTP/2 senders to catch GOAWAY/RST states before forwarding requests
+- Return connection age from the HTTP/2 pool checkout path and log warnings for older pooled connections
+- Evict pooled HTTP/2 senders when they are closed, exceed max age, fail readiness validation, or time out during readiness checks
+
+## 2026-03-16 - 25.11.9 - fix(rustproxy-routing)
+reduce hot-path allocations in routing, metrics, and proxy protocol handling
+
+- skip HTTP header map construction unless a route on the current port uses header matching
+- reuse computed client IP strings during HTTP route matching to avoid redundant allocations
+- optimize per-route and per-IP metric updates with get-first lookups to avoid unnecessary String creation on existing entries
+- replace heap-allocated PROXY protocol peek and discard buffers with stack-allocated buffers in the TCP listener
+- improve domain matcher case-insensitive wildcard checks while preserving glob fallback behavior
+
+## 2026-03-16 - 25.11.8 - fix(rustproxy-http)
+prevent premature idle timeouts during streamed HTTP responses and ensure TLS close_notify is sent on dropped connections
+
+- track active streaming response bodies so the HTTP idle watchdog does not close connections mid-transfer
+- add a ShutdownOnDrop wrapper for TLS-terminated HTTP connections to send shutdown on drop and avoid improperly terminated TLS sessions
+- apply the shutdown wrapper in passthrough TLS terminate and terminate+reencrypt HTTP handling
+
+## 2026-03-16 - 25.11.7 - fix(rustproxy)
+prevent TLS route reload certificate mismatches and tighten passthrough connection handling
+
+- Load updated TLS configs before swapping the route manager so newly visible routes always have their certificates available.
+- Add timeouts when peeking initial decrypted data after TLS handshake to avoid leaked idle connections.
+- Raise dropped, blocked, unmatched, and errored passthrough connection events from debug to warn for better operational visibility.
+
+## 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
+
+- Adds domain context to backend TCP/TLS connect, handshake, request failure, retry, and fallback log entries in the Rust HTTP proxy service.
+- Propagates the resolved host/domain through H1, H2, pooled, and fallback forwarding paths so backend-level diagnostics can be correlated with the original request domain.
+
+## 2026-03-12 - 25.10.2 - fix(repo)
+no code changes to release
+
+
+## 2026-03-12 - 25.10.1 - fix(repo)
+no changes to commit
+
+
+## 2026-03-12 - 25.10.0 - feat(metrics)
+add per-backend connection, error, protocol, and pool metrics with stale backend pruning
+
+- tracks backend connection lifecycle, connect timing, protocol detection, pool hit/miss rates, handshake/request errors, and h2 fallback failures in Rust metrics
+- exposes backend metrics through the TypeScript metrics adapter with backend listings, protocol lookup, and top error summaries
+- prunes backend metrics for backends no longer referenced by active routes, including preserved-port targets expanded across listening ports
+
+## 2026-03-11 - 25.9.3 - fix(rustproxy-http)
+Evict stale HTTP/2 pooled senders and retry bodyless requests with fresh backend connections to avoid 502s
+
+- Introduce MAX_H2_AGE (120s) and evict HTTP/2 senders older than this or closed
+- Check MAX_H2_AGE on checkout and during background eviction to prevent reuse of stale h2 connections
+- Add connection_pool.remove_h2() to explicitly remove dead H2 senders from the pool
+- When a pooled H2 request returns a 502 and the original request had an empty body, retry using a fresh H2 connection (retry_h2_with_fresh_connection)
+- On H2 auto-detect failures, retry as HTTP/1.1 for bodyless requests via forward_h1_empty_body; return 502 for requests with bodies
+- Evict dead H2 senders on backend request failures in reconnect_backend so subsequent attempts create fresh connections
+
+## 2026-03-08 - 25.9.2 - fix(protocol-cache)
+Include requested_host in protocol detection cache key to avoid cache oscillation when multiple frontend domains share the same backend
+
+- Add ProtocolCacheKey.requested_host: Option<String> to distinguish cache entries by incoming request Host/:authority
+- Update protocol cache lookups/inserts in proxy_service to populate requested_host
+- Enhance debug logging to show requested_host on cache hits
+- Fixes repeated ALPN probing / cache oscillation when different frontend domains share a backend with differing HTTP/2 support
+
+## 2026-03-03 - 25.9.1 - fix(rustproxy)
+Cancel connections for routes removed/disabled by adding per-route cancellation tokens and make RouteManager swappable (ArcSwap) for runtime updates
+
+- Add per-route CancellationToken map (DashMap) to TcpListenerManager and call token.cancel() when routes are removed (invalidate_removed_routes)
+- Propagate Arc<ArcSwap<RouteManager>> into HttpProxyService and passthrough listener so the route manager can be hot-swapped without restarting listeners
+- Use per-route child cancellation tokens in accept/connection handling and forwarders to terminate existing connections when a route is removed
+- Prune HTTP proxy caches and retain/cleanup per-route tokens when routes are active/removed
+- Update test.test.sni-requirement.node.ts to allocate unique free ports via findFreePorts to avoid port conflicts during tests
+
+## 2026-03-03 - 25.9.0 - feat(rustproxy-http)
+add HTTP/2 auto-detection via ALPN with TTL-backed protocol cache and h1-only/h2 ALPN client configs
+
+- Add protocol_cache module: bounded, TTL-based cache (5min TTL), max entries (4096), background cleanup task and clear() to discard stale detections.
+- Introduce BackendProtocol::Auto and expose 'auto' in TypeScript route types to allow ALPN-based protocol auto-detection.
+- Add build_tls_acceptor_h1_only() to create a TLS acceptor that advertises only http/1.1 (used for backends/tests that speak plain HTTP/1.1).
+- Add shared_backend_tls_config_alpn() and default_backend_tls_config_with_alpn() to provide client TLS configs advertising h2+http/1.1 for auto-detection.
+- Wire backend_tls_config_alpn and protocol_cache into proxy_service, tcp_listener and passthrough paths; add set_backend_tls_config_alpn() and prune protocol_cache on route updates.
+- Update passthrough tests to use h1-only acceptor to avoid false HTTP/2 detection when backends speak plain HTTP/1.1.
+- Include reconnection/fallback handling and ensure ALPN-enabled client config is used for auto-detection mode.
+
+## 2026-02-26 - 25.8.5 - fix(release)
+bump patch version (no source changes)
+
+- No changes detected in git diff
+- Current version: 25.8.4
+- Recommend patch bump to 25.8.5 to record release without code changes
+
+## 2026-02-26 - 25.8.4 - fix(proxy)
+adjust default proxy timeouts and keep-alive behavior to shorter, more consistent values
+
+- Increase connection timeout default from 30,000ms to 60,000ms (30s -> 60s).
+- Reduce socket timeout default from 3,600,000ms to 60,000ms (1h -> 60s).
+- Reduce max connection lifetime default from 86,400,000ms to 3,600,000ms (24h -> 1h).
+- Change inactivity timeout default from 14,400,000ms to 75,000ms (4h -> 75s).
+- Update keep-alive defaults: keepAliveTreatment 'extended' -> 'standard', keepAliveInactivityMultiplier 6 -> 4, extendedKeepAliveLifetime 604800000 -> 3,600,000ms (7d -> 1h).
+- Apply these consistent default values across Rust crates (rustproxy-config, rustproxy-passthrough) and the TypeScript smart-proxy implementation.
+- Update unit test expectations to match the new defaults.
+
 ## 2026-02-26 - 25.8.3 - fix(smartproxy)
 no code or dependency changes detected; no version bump required
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@push.rocks/smartproxy",
-  "version": "25.8.3",
+  "version": "25.11.14",
   "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-config/src/proxy_options.rs

@@ -298,7 +298,7 @@ impl RustProxyOptions {
 
     /// Get the effective connection timeout in milliseconds.
     pub fn effective_connection_timeout(&self) -> u64 {
-        self.connection_timeout.unwrap_or(30_000)
+        self.connection_timeout.unwrap_or(60_000)
     }
 
     /// Get the effective initial data timeout in milliseconds.
@@ -308,12 +308,12 @@ impl RustProxyOptions {
 
     /// Get the effective socket timeout in milliseconds.
     pub fn effective_socket_timeout(&self) -> u64 {
-        self.socket_timeout.unwrap_or(3_600_000)
+        self.socket_timeout.unwrap_or(60_000)
     }
 
     /// Get the effective max connection lifetime in milliseconds.
     pub fn effective_max_connection_lifetime(&self) -> u64 {
-        self.max_connection_lifetime.unwrap_or(86_400_000)
+        self.max_connection_lifetime.unwrap_or(3_600_000)
     }
 
     /// Get all unique ports that routes listen on.
@@ -377,10 +377,10 @@ mod tests {
     #[test]
     fn test_default_timeouts() {
         let options = RustProxyOptions::default();
-        assert_eq!(options.effective_connection_timeout(), 30_000);
+        assert_eq!(options.effective_connection_timeout(), 60_000);
         assert_eq!(options.effective_initial_data_timeout(), 60_000);
-        assert_eq!(options.effective_socket_timeout(), 3_600_000);
-        assert_eq!(options.effective_max_connection_lifetime(), 86_400_000);
+        assert_eq!(options.effective_socket_timeout(), 60_000);
+        assert_eq!(options.effective_max_connection_lifetime(), 3_600_000);
     }
 
     #[test]

rust/crates/rustproxy-config/src/route_types.rs

@@ -367,6 +367,7 @@ pub struct NfTablesOptions {
 pub enum BackendProtocol {
     Http1,
     Http2,
+    Auto,
 }
 
 /// Action options.

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

@@ -10,7 +10,7 @@ use bytes::Bytes;
 use dashmap::DashMap;
 use http_body_util::combinators::BoxBody;
 use hyper::client::conn::{http1, http2};
-use tracing::debug;
+// No per-request logging in the pool — only log on actual failures (in proxy_service.rs)
 
 /// Maximum idle connections per backend key.
 const MAX_IDLE_PER_KEY: usize = 16;
@@ -18,6 +18,10 @@ const MAX_IDLE_PER_KEY: usize = 16;
 const IDLE_TIMEOUT: Duration = Duration::from_secs(90);
 /// Background eviction interval.
 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.
 #[derive(Clone, Debug, Hash, Eq, PartialEq)]
@@ -37,7 +41,6 @@ struct IdleH1 {
 /// A pooled HTTP/2 sender (multiplexed, Clone-able).
 struct PooledH2 {
     sender: http2::SendRequest<BoxBody<Bytes, hyper::Error>>,
-    #[allow(dead_code)] // Reserved for future age-based eviction
     created_at: Instant,
 }
 
@@ -79,7 +82,7 @@ impl ConnectionPool {
         while let Some(idle) = idles.pop() {
             // Check if the connection is still alive and ready
             if idle.idle_since.elapsed() < IDLE_TIMEOUT && idle.sender.is_ready() && !idle.sender.is_closed() {
-                debug!("Pool hit (h1): {}:{}", key.host, key.port);
+                // H1 pool hit — no logging on hot path
                 return Some(idle.sender);
             }
             // Stale or closed — drop it
@@ -112,24 +115,29 @@ impl ConnectionPool {
 
     /// Try to get a cloned HTTP/2 sender for the given key.
     /// HTTP/2 senders are Clone-able (multiplexed), so we clone rather than remove.
-    pub fn checkout_h2(&self, key: &PoolKey) -> Option<http2::SendRequest<BoxBody<Bytes, hyper::Error>>> {
+    pub fn checkout_h2(&self, key: &PoolKey) -> Option<(http2::SendRequest<BoxBody<Bytes, hyper::Error>>, Duration)> {
         let entry = self.h2_pool.get(key)?;
         let pooled = entry.value();
+        let age = pooled.created_at.elapsed();
 
-        // Check if the h2 connection is still alive
-        if pooled.sender.is_closed() {
+        if pooled.sender.is_closed() || age >= MAX_H2_AGE {
             drop(entry);
             self.h2_pool.remove(key);
             return None;
         }
 
         if pooled.sender.is_ready() {
-            debug!("Pool hit (h2): {}:{}", key.host, key.port);
-            return Some(pooled.sender.clone());
+            return Some((pooled.sender.clone(), age));
         }
         None
     }
 
+    /// Remove a dead HTTP/2 sender from the pool.
+    /// Called when `send_request` fails to prevent subsequent requests from reusing the stale sender.
+    pub fn remove_h2(&self, key: &PoolKey) {
+        self.h2_pool.remove(key);
+    }
+
     /// Register an HTTP/2 sender in the pool. Since h2 is multiplexed,
     /// only one sender per key is stored (it's Clone-able).
     pub fn register_h2(&self, key: PoolKey, sender: http2::SendRequest<BoxBody<Bytes, hyper::Error>>) {
@@ -165,10 +173,10 @@ impl ConnectionPool {
                 h1_pool.remove(&key);
             }
 
-            // Evict dead H2 connections
+            // Evict dead or aged-out H2 connections
             let mut dead_h2 = Vec::new();
             for entry in h2_pool.iter() {
-                if entry.value().sender.is_closed() {
+                if entry.value().sender.is_closed() || entry.value().created_at.elapsed() >= MAX_H2_AGE {
                     dead_h2.push(entry.key().clone());
                 }
             }

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

@@ -11,20 +11,26 @@ 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>,
+    /// Optional active-request counter. When set, CountingBody increments on creation
+    /// and decrements on Drop, keeping the HTTP idle watchdog aware that a response
+    /// body is still streaming (even after the request handler has returned).
+    active_requests: Option<Arc<AtomicU64>>,
 }
 
 /// Which direction the bytes flow.
@@ -47,42 +53,46 @@ 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,
+            active_requests: None,
         }
     }
 
-    /// Report accumulated bytes to the metrics collector.
-    fn report(&mut self) {
-        if self.reported {
-            return;
-        }
-        self.reported = true;
+    /// 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
+    }
 
-        let bytes = self.counted_bytes.load(Ordering::Relaxed);
-        if bytes == 0 {
-            return;
-        }
+    /// Set the active-request counter for the HTTP idle watchdog.
+    /// CountingBody increments on creation and decrements on Drop, ensuring the
+    /// idle watchdog sees an "active request" while the response body streams.
+    pub fn with_active_requests(mut self, counter: Arc<AtomicU64>) -> Self {
+        counter.fetch_add(1, Ordering::Relaxed);
+        self.active_requests = Some(counter);
+        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 +112,18 @@ 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,
         }
     }
@@ -124,3 +136,13 @@ where
         self.inner.size_hint()
     }
 }
+
+impl<B> Drop for CountingBody<B> {
+    fn drop(&mut self) {
+        // Decrement the active-request counter so the HTTP idle watchdog
+        // knows this response body is no longer streaming.
+        if let Some(ref counter) = self.active_requests {
+            counter.fetch_sub(1, Ordering::Relaxed);
+        }
+    }
+}

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

@@ -5,9 +5,11 @@
 
 pub mod connection_pool;
 pub mod counting_body;
+pub mod protocol_cache;
 pub mod proxy_service;
 pub mod request_filter;
 pub mod response_filter;
+pub mod shutdown_on_drop;
 pub mod template;
 pub mod upstream_selector;
 

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

@@ -0,0 +1,140 @@
+//! Bounded, TTL-based protocol detection cache for HTTP/2 auto-detection.
+//!
+//! Caches the ALPN-negotiated protocol (H1 or H2) per backend endpoint and requested
+//! domain (host:port + requested_host). This prevents cache oscillation when multiple
+//! frontend domains share the same backend but differ in HTTP/2 support.
+
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use dashmap::DashMap;
+use tracing::debug;
+
+/// TTL for cached protocol detection results.
+/// After this duration, the next request will re-probe the backend.
+const PROTOCOL_CACHE_TTL: Duration = Duration::from_secs(300); // 5 minutes
+
+/// Maximum number of entries in the protocol cache.
+/// Prevents unbounded growth when backends come and go.
+const PROTOCOL_CACHE_MAX_ENTRIES: usize = 4096;
+
+/// Background cleanup interval for the protocol cache.
+const PROTOCOL_CACHE_CLEANUP_INTERVAL: Duration = Duration::from_secs(60);
+
+/// Detected backend protocol.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum DetectedProtocol {
+    H1,
+    H2,
+}
+
+/// Key for the protocol cache: (host, port, requested_host).
+#[derive(Clone, Debug, Hash, Eq, PartialEq)]
+pub struct ProtocolCacheKey {
+    pub host: String,
+    pub port: u16,
+    /// The incoming request's domain (Host header / :authority).
+    /// Distinguishes protocol detection when multiple domains share the same backend.
+    pub requested_host: Option<String>,
+}
+
+/// A cached protocol detection result with a timestamp.
+struct CachedEntry {
+    protocol: DetectedProtocol,
+    detected_at: Instant,
+}
+
+/// Bounded, TTL-based protocol detection cache.
+///
+/// Memory safety guarantees:
+/// - Hard cap at `PROTOCOL_CACHE_MAX_ENTRIES` — cannot grow unboundedly.
+/// - TTL expiry — stale entries naturally age out on lookup.
+/// - Background cleanup task — proactively removes expired entries every 60s.
+/// - `clear()` — called on route updates to discard stale detections.
+/// - `Drop` — aborts the background task to prevent dangling tokio tasks.
+pub struct ProtocolCache {
+    cache: Arc<DashMap<ProtocolCacheKey, CachedEntry>>,
+    cleanup_handle: Option<tokio::task::JoinHandle<()>>,
+}
+
+impl ProtocolCache {
+    /// Create a new protocol cache and start the background cleanup task.
+    pub fn new() -> Self {
+        let cache: Arc<DashMap<ProtocolCacheKey, CachedEntry>> = Arc::new(DashMap::new());
+        let cache_clone = Arc::clone(&cache);
+        let cleanup_handle = tokio::spawn(async move {
+            Self::cleanup_loop(cache_clone).await;
+        });
+
+        Self {
+            cache,
+            cleanup_handle: Some(cleanup_handle),
+        }
+    }
+
+    /// Look up the cached protocol for a backend endpoint.
+    /// Returns `None` if not cached or expired (caller should probe via ALPN).
+    pub fn get(&self, key: &ProtocolCacheKey) -> Option<DetectedProtocol> {
+        let entry = self.cache.get(key)?;
+        if entry.detected_at.elapsed() < PROTOCOL_CACHE_TTL {
+            debug!("Protocol cache hit: {:?} for {}:{} (requested: {:?})", entry.protocol, key.host, key.port, key.requested_host);
+            Some(entry.protocol)
+        } else {
+            // Expired — remove and return None to trigger re-probe
+            drop(entry); // release DashMap ref before remove
+            self.cache.remove(key);
+            None
+        }
+    }
+
+    /// Insert a detected protocol into the cache.
+    /// If the cache is at capacity, evict the oldest entry first.
+    pub fn insert(&self, key: ProtocolCacheKey, protocol: DetectedProtocol) {
+        if self.cache.len() >= PROTOCOL_CACHE_MAX_ENTRIES && !self.cache.contains_key(&key) {
+            // Evict the oldest entry to stay within bounds
+            let oldest = self.cache.iter()
+                .min_by_key(|entry| entry.value().detected_at)
+                .map(|entry| entry.key().clone());
+            if let Some(oldest_key) = oldest {
+                self.cache.remove(&oldest_key);
+            }
+        }
+        self.cache.insert(key, CachedEntry {
+            protocol,
+            detected_at: Instant::now(),
+        });
+    }
+
+    /// Clear all entries. Called on route updates to discard stale detections.
+    pub fn clear(&self) {
+        self.cache.clear();
+    }
+
+    /// Background cleanup loop — removes expired entries every `PROTOCOL_CACHE_CLEANUP_INTERVAL`.
+    async fn cleanup_loop(cache: Arc<DashMap<ProtocolCacheKey, CachedEntry>>) {
+        let mut interval = tokio::time::interval(PROTOCOL_CACHE_CLEANUP_INTERVAL);
+        loop {
+            interval.tick().await;
+
+            let expired: Vec<ProtocolCacheKey> = cache.iter()
+                .filter(|entry| entry.value().detected_at.elapsed() >= PROTOCOL_CACHE_TTL)
+                .map(|entry| entry.key().clone())
+                .collect();
+
+            if !expired.is_empty() {
+                debug!("Protocol cache cleanup: removing {} expired entries", expired.len());
+                for key in expired {
+                    cache.remove(&key);
+                }
+            }
+        }
+    }
+}
+
+impl Drop for ProtocolCache {
+    fn drop(&mut self) {
+        if let Some(handle) = self.cleanup_handle.take() {
+            handle.abort();
+        }
+    }
+}

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

@@ -0,0 +1,102 @@
+//! Wrapper that ensures TLS close_notify is sent when the stream is dropped.
+//!
+//! When hyper drops an HTTP connection (backend error, timeout, normal H2 close),
+//! the underlying TLS stream is dropped WITHOUT `shutdown()`. tokio-rustls cannot
+//! send `close_notify` in Drop (requires async). This wrapper tracks whether
+//! `poll_shutdown` was called and, if not, spawns a background task to send it.
+
+use std::io;
+use std::pin::Pin;
+use std::task::{Context, Poll};
+
+use tokio::io::{AsyncRead, AsyncWrite, ReadBuf};
+
+/// Wraps an AsyncRead+AsyncWrite stream and ensures `shutdown()` is called when
+/// dropped, even if the caller (e.g. hyper) doesn't explicitly shut down.
+///
+/// This guarantees TLS `close_notify` is sent for TLS-wrapped streams, preventing
+/// "GnuTLS recv error (-110): The TLS connection was non-properly terminated" errors.
+pub struct ShutdownOnDrop<S: AsyncRead + AsyncWrite + Unpin + Send + 'static> {
+    inner: Option<S>,
+    shutdown_called: bool,
+}
+
+impl<S: AsyncRead + AsyncWrite + Unpin + Send + 'static> ShutdownOnDrop<S> {
+    /// Create a new wrapper around the given stream.
+    pub fn new(stream: S) -> Self {
+        Self {
+            inner: Some(stream),
+            shutdown_called: false,
+        }
+    }
+}
+
+impl<S: AsyncRead + AsyncWrite + Unpin + Send + 'static> AsyncRead for ShutdownOnDrop<S> {
+    fn poll_read(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &mut ReadBuf<'_>,
+    ) -> Poll<io::Result<()>> {
+        Pin::new(self.get_mut().inner.as_mut().unwrap()).poll_read(cx, buf)
+    }
+}
+
+impl<S: AsyncRead + AsyncWrite + Unpin + Send + 'static> AsyncWrite for ShutdownOnDrop<S> {
+    fn poll_write(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<io::Result<usize>> {
+        Pin::new(self.get_mut().inner.as_mut().unwrap()).poll_write(cx, buf)
+    }
+
+    fn poll_write_vectored(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        bufs: &[io::IoSlice<'_>],
+    ) -> Poll<io::Result<usize>> {
+        Pin::new(self.get_mut().inner.as_mut().unwrap()).poll_write_vectored(cx, bufs)
+    }
+
+    fn is_write_vectored(&self) -> bool {
+        self.inner.as_ref().unwrap().is_write_vectored()
+    }
+
+    fn poll_flush(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+    ) -> Poll<io::Result<()>> {
+        Pin::new(self.get_mut().inner.as_mut().unwrap()).poll_flush(cx)
+    }
+
+    fn poll_shutdown(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+    ) -> Poll<io::Result<()>> {
+        let this = self.get_mut();
+        let result = Pin::new(this.inner.as_mut().unwrap()).poll_shutdown(cx);
+        if result.is_ready() {
+            this.shutdown_called = true;
+        }
+        result
+    }
+}
+
+impl<S: AsyncRead + AsyncWrite + Unpin + Send + 'static> Drop for ShutdownOnDrop<S> {
+    fn drop(&mut self) {
+        // If shutdown was already called (hyper closed properly), nothing to do.
+        // If not (hyper dropped without shutdown — e.g. H2 close, error, timeout),
+        // spawn a background task to send close_notify / TCP FIN.
+        if !self.shutdown_called {
+            if let Some(mut stream) = self.inner.take() {
+                tokio::spawn(async move {
+                    let _ = tokio::time::timeout(
+                        std::time::Duration::from_secs(2),
+                        tokio::io::AsyncWriteExt::shutdown(&mut stream),
+                    ).await;
+                    // stream is dropped here — all resources freed
+                });
+            }
+        }
+    }
+}

rust/crates/rustproxy-metrics/src/collector.rs

@@ -3,6 +3,7 @@ use serde::{Deserialize, Serialize};
 use std::collections::HashSet;
 use std::sync::atomic::{AtomicU64, Ordering};
 use std::sync::Mutex;
+use std::time::Duration;
 
 use crate::throughput::{ThroughputSample, ThroughputTracker};
 
@@ -20,6 +21,7 @@ pub struct Metrics {
     pub throughput_recent_out_bytes_per_sec: u64,
     pub routes: std::collections::HashMap<String, RouteMetrics>,
     pub ips: std::collections::HashMap<String, IpMetrics>,
+    pub backends: std::collections::HashMap<String, BackendMetrics>,
     pub throughput_history: Vec<ThroughputSample>,
     pub total_http_requests: u64,
     pub http_requests_per_sec: u64,
@@ -52,6 +54,23 @@ pub struct IpMetrics {
     pub throughput_out_bytes_per_sec: u64,
 }
 
+/// Per-backend metrics (keyed by "host:port").
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct BackendMetrics {
+    pub active_connections: u64,
+    pub total_connections: u64,
+    pub protocol: String,
+    pub connect_errors: u64,
+    pub handshake_errors: u64,
+    pub request_errors: u64,
+    pub total_connect_time_us: u64,
+    pub connect_count: u64,
+    pub pool_hits: u64,
+    pub pool_misses: u64,
+    pub h2_failures: u64,
+}
+
 /// Statistics snapshot.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
@@ -69,6 +88,9 @@ const DEFAULT_RETENTION_SECONDS: usize = 3600;
 /// Maximum number of IPs to include in a snapshot (top by active connections).
 const MAX_IPS_IN_SNAPSHOT: usize = 100;
 
+/// Maximum number of backends to include in a snapshot (top by total connections).
+const MAX_BACKENDS_IN_SNAPSHOT: usize = 100;
+
 /// Metrics collector tracking connections and throughput.
 ///
 /// Design: The hot path (`record_bytes`) is entirely lock-free — it only touches
@@ -96,6 +118,19 @@ pub struct MetricsCollector {
     ip_pending_tp: DashMap<String, (AtomicU64, AtomicU64)>,
     ip_throughput: DashMap<String, Mutex<ThroughputTracker>>,
 
+    // ── Per-backend tracking (keyed by "host:port") ──
+    backend_active: DashMap<String, AtomicU64>,
+    backend_total: DashMap<String, AtomicU64>,
+    backend_protocol: DashMap<String, String>,
+    backend_connect_errors: DashMap<String, AtomicU64>,
+    backend_handshake_errors: DashMap<String, AtomicU64>,
+    backend_request_errors: DashMap<String, AtomicU64>,
+    backend_connect_time_us: DashMap<String, AtomicU64>,
+    backend_connect_count: DashMap<String, AtomicU64>,
+    backend_pool_hits: DashMap<String, AtomicU64>,
+    backend_pool_misses: DashMap<String, AtomicU64>,
+    backend_h2_failures: DashMap<String, AtomicU64>,
+
     // ── HTTP request tracking ──
     total_http_requests: AtomicU64,
     pending_http_requests: AtomicU64,
@@ -134,6 +169,17 @@ impl MetricsCollector {
             ip_bytes_out: DashMap::new(),
             ip_pending_tp: DashMap::new(),
             ip_throughput: DashMap::new(),
+            backend_active: DashMap::new(),
+            backend_total: DashMap::new(),
+            backend_protocol: DashMap::new(),
+            backend_connect_errors: DashMap::new(),
+            backend_handshake_errors: DashMap::new(),
+            backend_request_errors: DashMap::new(),
+            backend_connect_time_us: DashMap::new(),
+            backend_connect_count: DashMap::new(),
+            backend_pool_hits: DashMap::new(),
+            backend_pool_misses: DashMap::new(),
+            backend_h2_failures: DashMap::new(),
             total_http_requests: AtomicU64::new(0),
             pending_http_requests: AtomicU64::new(0),
             http_request_throughput: Mutex::new(ThroughputTracker::new(retention_seconds)),
@@ -220,44 +266,67 @@ impl MetricsCollector {
         self.global_pending_tp_in.fetch_add(bytes_in, Ordering::Relaxed);
         self.global_pending_tp_out.fetch_add(bytes_out, Ordering::Relaxed);
 
+        // Per-route tracking: use get() first (zero-alloc fast path for existing entries),
+        // fall back to entry() with to_string() only on the rare first-chunk miss.
         if let Some(route_id) = route_id {
-            self.route_bytes_in
-                .entry(route_id.to_string())
-                .or_insert_with(|| AtomicU64::new(0))
-                .fetch_add(bytes_in, Ordering::Relaxed);
-            self.route_bytes_out
-                .entry(route_id.to_string())
-                .or_insert_with(|| AtomicU64::new(0))
-                .fetch_add(bytes_out, Ordering::Relaxed);
+            if let Some(counter) = self.route_bytes_in.get(route_id) {
+                counter.fetch_add(bytes_in, Ordering::Relaxed);
+            } else {
+                self.route_bytes_in.entry(route_id.to_string())
+                    .or_insert_with(|| AtomicU64::new(0))
+                    .fetch_add(bytes_in, Ordering::Relaxed);
+            }
+            if let Some(counter) = self.route_bytes_out.get(route_id) {
+                counter.fetch_add(bytes_out, Ordering::Relaxed);
+            } else {
+                self.route_bytes_out.entry(route_id.to_string())
+                    .or_insert_with(|| AtomicU64::new(0))
+                    .fetch_add(bytes_out, Ordering::Relaxed);
+            }
 
             // Accumulate into per-route pending throughput counters (lock-free)
-            let entry = self.route_pending_tp
-                .entry(route_id.to_string())
-                .or_insert_with(|| (AtomicU64::new(0), AtomicU64::new(0)));
-            entry.0.fetch_add(bytes_in, Ordering::Relaxed);
-            entry.1.fetch_add(bytes_out, Ordering::Relaxed);
+            if let Some(entry) = self.route_pending_tp.get(route_id) {
+                entry.0.fetch_add(bytes_in, Ordering::Relaxed);
+                entry.1.fetch_add(bytes_out, Ordering::Relaxed);
+            } else {
+                let entry = self.route_pending_tp.entry(route_id.to_string())
+                    .or_insert_with(|| (AtomicU64::new(0), AtomicU64::new(0)));
+                entry.0.fetch_add(bytes_in, Ordering::Relaxed);
+                entry.1.fetch_add(bytes_out, Ordering::Relaxed);
+            }
         }
 
+        // Per-IP tracking: same get()-first pattern to avoid String allocation on hot path.
         if let Some(ip) = source_ip {
             // Only record per-IP stats if the IP still has active connections.
             // This prevents orphaned entries when record_bytes races with
             // connection_closed (which evicts all per-IP data on last close).
             if self.ip_connections.contains_key(ip) {
-                self.ip_bytes_in
-                    .entry(ip.to_string())
-                    .or_insert_with(|| AtomicU64::new(0))
-                    .fetch_add(bytes_in, Ordering::Relaxed);
-                self.ip_bytes_out
-                    .entry(ip.to_string())
-                    .or_insert_with(|| AtomicU64::new(0))
-                    .fetch_add(bytes_out, Ordering::Relaxed);
+                if let Some(counter) = self.ip_bytes_in.get(ip) {
+                    counter.fetch_add(bytes_in, Ordering::Relaxed);
+                } else {
+                    self.ip_bytes_in.entry(ip.to_string())
+                        .or_insert_with(|| AtomicU64::new(0))
+                        .fetch_add(bytes_in, Ordering::Relaxed);
+                }
+                if let Some(counter) = self.ip_bytes_out.get(ip) {
+                    counter.fetch_add(bytes_out, Ordering::Relaxed);
+                } else {
+                    self.ip_bytes_out.entry(ip.to_string())
+                        .or_insert_with(|| AtomicU64::new(0))
+                        .fetch_add(bytes_out, Ordering::Relaxed);
+                }
 
                 // Accumulate into per-IP pending throughput counters (lock-free)
-                let entry = self.ip_pending_tp
-                    .entry(ip.to_string())
-                    .or_insert_with(|| (AtomicU64::new(0), AtomicU64::new(0)));
-                entry.0.fetch_add(bytes_in, Ordering::Relaxed);
-                entry.1.fetch_add(bytes_out, Ordering::Relaxed);
+                if let Some(entry) = self.ip_pending_tp.get(ip) {
+                    entry.0.fetch_add(bytes_in, Ordering::Relaxed);
+                    entry.1.fetch_add(bytes_out, Ordering::Relaxed);
+                } else {
+                    let entry = self.ip_pending_tp.entry(ip.to_string())
+                        .or_insert_with(|| (AtomicU64::new(0), AtomicU64::new(0)));
+                    entry.0.fetch_add(bytes_in, Ordering::Relaxed);
+                    entry.1.fetch_add(bytes_out, Ordering::Relaxed);
+                }
             }
         }
     }
@@ -268,6 +337,113 @@ impl MetricsCollector {
         self.pending_http_requests.fetch_add(1, Ordering::Relaxed);
     }
 
+    // ── Per-backend recording methods ──
+
+    /// Record a successful backend connection with its connect duration.
+    pub fn backend_connection_opened(&self, key: &str, connect_time: Duration) {
+        self.backend_active
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+        self.backend_total
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+        self.backend_connect_time_us
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(connect_time.as_micros() as u64, Ordering::Relaxed);
+        self.backend_connect_count
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Record a backend connection closing.
+    pub fn backend_connection_closed(&self, key: &str) {
+        if let Some(counter) = self.backend_active.get(key) {
+            let val = counter.load(Ordering::Relaxed);
+            if val > 0 {
+                counter.fetch_sub(1, Ordering::Relaxed);
+            }
+        }
+    }
+
+    /// Record a backend connect error (TCP or TLS connect failure/timeout).
+    pub fn backend_connect_error(&self, key: &str) {
+        self.backend_connect_errors
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Record a backend handshake error (H1 or H2 handshake failure).
+    pub fn backend_handshake_error(&self, key: &str) {
+        self.backend_handshake_errors
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Record a backend request error (send_request failure).
+    pub fn backend_request_error(&self, key: &str) {
+        self.backend_request_errors
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Record a connection pool hit for a backend.
+    pub fn backend_pool_hit(&self, key: &str) {
+        self.backend_pool_hits
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Record a connection pool miss for a backend.
+    pub fn backend_pool_miss(&self, key: &str) {
+        self.backend_pool_misses
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Record an H2 failure (h2 attempted but fell back to h1).
+    pub fn backend_h2_failure(&self, key: &str) {
+        self.backend_h2_failures
+            .entry(key.to_string())
+            .or_insert_with(|| AtomicU64::new(0))
+            .fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Set the protocol in use for a backend ("h1" or "h2").
+    pub fn set_backend_protocol(&self, key: &str, protocol: &str) {
+        self.backend_protocol
+            .entry(key.to_string())
+            .and_modify(|v| {
+                if v != protocol {
+                    *v = protocol.to_string();
+                }
+            })
+            .or_insert_with(|| protocol.to_string());
+    }
+
+    /// Remove per-backend metrics for backends no longer in any route target.
+    pub fn retain_backends(&self, active_backends: &HashSet<String>) {
+        self.backend_active.retain(|k, _| active_backends.contains(k));
+        self.backend_total.retain(|k, _| active_backends.contains(k));
+        self.backend_protocol.retain(|k, _| active_backends.contains(k));
+        self.backend_connect_errors.retain(|k, _| active_backends.contains(k));
+        self.backend_handshake_errors.retain(|k, _| active_backends.contains(k));
+        self.backend_request_errors.retain(|k, _| active_backends.contains(k));
+        self.backend_connect_time_us.retain(|k, _| active_backends.contains(k));
+        self.backend_connect_count.retain(|k, _| active_backends.contains(k));
+        self.backend_pool_hits.retain(|k, _| active_backends.contains(k));
+        self.backend_pool_misses.retain(|k, _| active_backends.contains(k));
+        self.backend_h2_failures.retain(|k, _| active_backends.contains(k));
+    }
+
     /// Take a throughput sample on all trackers (cold path, call at 1Hz or configured interval).
     ///
     /// Drains the lock-free pending counters and feeds the accumulated bytes
@@ -488,6 +664,72 @@ impl MetricsCollector {
             });
         }
 
+        // Collect per-backend metrics, capped at top MAX_BACKENDS_IN_SNAPSHOT by total connections
+        let mut backend_entries: Vec<(String, BackendMetrics)> = Vec::new();
+        for entry in self.backend_total.iter() {
+            let key = entry.key().clone();
+            let total = entry.value().load(Ordering::Relaxed);
+            let active = self.backend_active
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let protocol = self.backend_protocol
+                .get(&key)
+                .map(|v| v.value().clone())
+                .unwrap_or_else(|| "unknown".to_string());
+            let connect_errors = self.backend_connect_errors
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let handshake_errors = self.backend_handshake_errors
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let request_errors = self.backend_request_errors
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let total_connect_time_us = self.backend_connect_time_us
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let connect_count = self.backend_connect_count
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let pool_hits = self.backend_pool_hits
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let pool_misses = self.backend_pool_misses
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+            let h2_failures = self.backend_h2_failures
+                .get(&key)
+                .map(|c| c.load(Ordering::Relaxed))
+                .unwrap_or(0);
+
+            backend_entries.push((key, BackendMetrics {
+                active_connections: active,
+                total_connections: total,
+                protocol,
+                connect_errors,
+                handshake_errors,
+                request_errors,
+                total_connect_time_us,
+                connect_count,
+                pool_hits,
+                pool_misses,
+                h2_failures,
+            }));
+        }
+        // Sort by total connections descending, then cap
+        backend_entries.sort_by(|a, b| b.1.total_connections.cmp(&a.1.total_connections));
+        backend_entries.truncate(MAX_BACKENDS_IN_SNAPSHOT);
+
+        let backends: std::collections::HashMap<String, BackendMetrics> = backend_entries.into_iter().collect();
+
         // HTTP request rates
         let (http_rps, http_rps_recent) = self.http_request_throughput
             .lock()
@@ -509,6 +751,7 @@ impl MetricsCollector {
             throughput_recent_out_bytes_per_sec: global_recent_out,
             routes,
             ips,
+            backends,
             throughput_history,
             total_http_requests: self.total_http_requests.load(Ordering::Relaxed),
             http_requests_per_sec: http_rps,
@@ -805,4 +1048,120 @@ mod tests {
         assert_eq!(snapshot.throughput_history[0].bytes_in, 100);
         assert_eq!(snapshot.throughput_history[4].bytes_in, 500);
     }
+
+    #[test]
+    fn test_backend_metrics_basic() {
+        let collector = MetricsCollector::new();
+        let key = "backend1:8080";
+
+        // Open connections with timing
+        collector.backend_connection_opened(key, Duration::from_millis(15));
+        collector.backend_connection_opened(key, Duration::from_millis(25));
+
+        assert_eq!(collector.backend_active.get(key).unwrap().load(Ordering::Relaxed), 2);
+        assert_eq!(collector.backend_total.get(key).unwrap().load(Ordering::Relaxed), 2);
+        assert_eq!(collector.backend_connect_count.get(key).unwrap().load(Ordering::Relaxed), 2);
+        // 15ms + 25ms = 40ms = 40_000us
+        assert_eq!(collector.backend_connect_time_us.get(key).unwrap().load(Ordering::Relaxed), 40_000);
+
+        // Close one
+        collector.backend_connection_closed(key);
+        assert_eq!(collector.backend_active.get(key).unwrap().load(Ordering::Relaxed), 1);
+        // total stays
+        assert_eq!(collector.backend_total.get(key).unwrap().load(Ordering::Relaxed), 2);
+
+        // Record errors
+        collector.backend_connect_error(key);
+        collector.backend_handshake_error(key);
+        collector.backend_request_error(key);
+        collector.backend_h2_failure(key);
+        collector.backend_pool_hit(key);
+        collector.backend_pool_hit(key);
+        collector.backend_pool_miss(key);
+
+        assert_eq!(collector.backend_connect_errors.get(key).unwrap().load(Ordering::Relaxed), 1);
+        assert_eq!(collector.backend_handshake_errors.get(key).unwrap().load(Ordering::Relaxed), 1);
+        assert_eq!(collector.backend_request_errors.get(key).unwrap().load(Ordering::Relaxed), 1);
+        assert_eq!(collector.backend_h2_failures.get(key).unwrap().load(Ordering::Relaxed), 1);
+        assert_eq!(collector.backend_pool_hits.get(key).unwrap().load(Ordering::Relaxed), 2);
+        assert_eq!(collector.backend_pool_misses.get(key).unwrap().load(Ordering::Relaxed), 1);
+
+        // Protocol
+        collector.set_backend_protocol(key, "h1");
+        assert_eq!(collector.backend_protocol.get(key).unwrap().value(), "h1");
+        collector.set_backend_protocol(key, "h2");
+        assert_eq!(collector.backend_protocol.get(key).unwrap().value(), "h2");
+    }
+
+    #[test]
+    fn test_backend_metrics_in_snapshot() {
+        let collector = MetricsCollector::new();
+
+        collector.backend_connection_opened("b1:443", Duration::from_millis(10));
+        collector.backend_connection_opened("b2:8080", Duration::from_millis(20));
+        collector.set_backend_protocol("b1:443", "h2");
+        collector.set_backend_protocol("b2:8080", "h1");
+        collector.backend_connect_error("b1:443");
+
+        let snapshot = collector.snapshot();
+        assert_eq!(snapshot.backends.len(), 2);
+
+        let b1 = snapshot.backends.get("b1:443").unwrap();
+        assert_eq!(b1.active_connections, 1);
+        assert_eq!(b1.total_connections, 1);
+        assert_eq!(b1.protocol, "h2");
+        assert_eq!(b1.connect_errors, 1);
+        assert_eq!(b1.total_connect_time_us, 10_000);
+        assert_eq!(b1.connect_count, 1);
+
+        let b2 = snapshot.backends.get("b2:8080").unwrap();
+        assert_eq!(b2.protocol, "h1");
+        assert_eq!(b2.connect_errors, 0);
+    }
+
+    #[test]
+    fn test_retain_backends_prunes_stale() {
+        let collector = MetricsCollector::new();
+
+        collector.backend_connection_opened("active:443", Duration::from_millis(5));
+        collector.backend_connection_opened("stale:8080", Duration::from_millis(10));
+        collector.set_backend_protocol("active:443", "h1");
+        collector.set_backend_protocol("stale:8080", "h2");
+        collector.backend_connect_error("stale:8080");
+
+        let active = HashSet::from(["active:443".to_string()]);
+        collector.retain_backends(&active);
+
+        // active:443 should still exist
+        assert!(collector.backend_total.get("active:443").is_some());
+        assert!(collector.backend_protocol.get("active:443").is_some());
+
+        // stale:8080 should be fully removed
+        assert!(collector.backend_active.get("stale:8080").is_none());
+        assert!(collector.backend_total.get("stale:8080").is_none());
+        assert!(collector.backend_protocol.get("stale:8080").is_none());
+        assert!(collector.backend_connect_errors.get("stale:8080").is_none());
+        assert!(collector.backend_connect_time_us.get("stale:8080").is_none());
+        assert!(collector.backend_connect_count.get("stale:8080").is_none());
+        assert!(collector.backend_pool_hits.get("stale:8080").is_none());
+        assert!(collector.backend_pool_misses.get("stale:8080").is_none());
+        assert!(collector.backend_h2_failures.get("stale:8080").is_none());
+    }
+
+    #[test]
+    fn test_backend_connection_closed_saturates() {
+        let collector = MetricsCollector::new();
+        let key = "b:80";
+
+        // Close without opening — should not underflow
+        collector.backend_connection_closed(key);
+        // No entry created
+        assert!(collector.backend_active.get(key).is_none());
+
+        // Open one, close two — should saturate at 0
+        collector.backend_connection_opened(key, Duration::from_millis(1));
+        collector.backend_connection_closed(key);
+        collector.backend_connection_closed(key);
+        assert_eq!(collector.backend_active.get(key).unwrap().load(Ordering::Relaxed), 0);
+    }
 }

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

@@ -1,6 +1,7 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 use arc_swap::ArcSwap;
+use dashmap::DashMap;
 use tokio::net::TcpListener;
 use tokio_rustls::TlsAcceptor;
 use tokio_util::sync::CancellationToken;
@@ -118,10 +119,10 @@ pub struct ConnectionConfig {
 impl Default for ConnectionConfig {
     fn default() -> Self {
         Self {
-            connection_timeout_ms: 30_000,
+            connection_timeout_ms: 60_000,
             initial_data_timeout_ms: 60_000,
-            socket_timeout_ms: 3_600_000,
-            max_connection_lifetime_ms: 86_400_000,
+            socket_timeout_ms: 60_000,
+            max_connection_lifetime_ms: 3_600_000,
             graceful_shutdown_timeout_ms: 30_000,
             max_connections_per_ip: None,
             connection_rate_limit_per_minute: None,
@@ -162,18 +163,28 @@ pub struct TcpListenerManager {
     socket_handler_relay: Arc<std::sync::RwLock<Option<String>>>,
     /// Global connection semaphore — limits total simultaneous connections.
     conn_semaphore: Arc<tokio::sync::Semaphore>,
+    /// Per-route cancellation tokens (child of cancel_token).
+    /// When a route is removed, its token is cancelled, terminating all connections on that route.
+    route_cancels: Arc<DashMap<String, CancellationToken>>,
 }
 
 impl TcpListenerManager {
     pub fn new(route_manager: Arc<RouteManager>) -> Self {
         let metrics = Arc::new(MetricsCollector::new());
         let conn_config = ConnectionConfig::default();
+        let route_manager_swap = Arc::new(ArcSwap::from(route_manager));
         let mut http_proxy_svc = HttpProxyService::with_connect_timeout(
-            Arc::clone(&route_manager),
+            Arc::clone(&route_manager_swap),
             Arc::clone(&metrics),
             std::time::Duration::from_millis(conn_config.connection_timeout_ms),
         );
         http_proxy_svc.set_backend_tls_config(tls_handler::shared_backend_tls_config());
+        http_proxy_svc.set_backend_tls_config_alpn(tls_handler::shared_backend_tls_config_alpn());
+        http_proxy_svc.set_connection_timeouts(
+            std::time::Duration::from_millis(conn_config.socket_timeout_ms),
+            std::time::Duration::from_millis(conn_config.socket_timeout_ms),
+            std::time::Duration::from_millis(conn_config.max_connection_lifetime_ms),
+        );
         let http_proxy = Arc::new(http_proxy_svc);
         let conn_tracker = Arc::new(ConnectionTracker::new(
             conn_config.max_connections_per_ip,
@@ -182,7 +193,7 @@ impl TcpListenerManager {
         let max_conns = conn_config.max_connections as usize;
         Self {
             listeners: HashMap::new(),
-            route_manager: Arc::new(ArcSwap::from(route_manager)),
+            route_manager: route_manager_swap,
             metrics,
             tls_configs: Arc::new(ArcSwap::from(Arc::new(HashMap::new()))),
             shared_tls_acceptor: Arc::new(ArcSwap::from(Arc::new(None))),
@@ -192,18 +203,26 @@ impl TcpListenerManager {
             cancel_token: CancellationToken::new(),
             socket_handler_relay: Arc::new(std::sync::RwLock::new(None)),
             conn_semaphore: Arc::new(tokio::sync::Semaphore::new(max_conns)),
+            route_cancels: Arc::new(DashMap::new()),
         }
     }
 
     /// Create with a metrics collector.
     pub fn with_metrics(route_manager: Arc<RouteManager>, metrics: Arc<MetricsCollector>) -> Self {
         let conn_config = ConnectionConfig::default();
+        let route_manager_swap = Arc::new(ArcSwap::from(route_manager));
         let mut http_proxy_svc = HttpProxyService::with_connect_timeout(
-            Arc::clone(&route_manager),
+            Arc::clone(&route_manager_swap),
             Arc::clone(&metrics),
             std::time::Duration::from_millis(conn_config.connection_timeout_ms),
         );
         http_proxy_svc.set_backend_tls_config(tls_handler::shared_backend_tls_config());
+        http_proxy_svc.set_backend_tls_config_alpn(tls_handler::shared_backend_tls_config_alpn());
+        http_proxy_svc.set_connection_timeouts(
+            std::time::Duration::from_millis(conn_config.socket_timeout_ms),
+            std::time::Duration::from_millis(conn_config.socket_timeout_ms),
+            std::time::Duration::from_millis(conn_config.max_connection_lifetime_ms),
+        );
         let http_proxy = Arc::new(http_proxy_svc);
         let conn_tracker = Arc::new(ConnectionTracker::new(
             conn_config.max_connections_per_ip,
@@ -212,7 +231,7 @@ impl TcpListenerManager {
         let max_conns = conn_config.max_connections as usize;
         Self {
             listeners: HashMap::new(),
-            route_manager: Arc::new(ArcSwap::from(route_manager)),
+            route_manager: route_manager_swap,
             metrics,
             tls_configs: Arc::new(ArcSwap::from(Arc::new(HashMap::new()))),
             shared_tls_acceptor: Arc::new(ArcSwap::from(Arc::new(None))),
@@ -222,6 +241,7 @@ impl TcpListenerManager {
             cancel_token: CancellationToken::new(),
             socket_handler_relay: Arc::new(std::sync::RwLock::new(None)),
             conn_semaphore: Arc::new(tokio::sync::Semaphore::new(max_conns)),
+            route_cancels: Arc::new(DashMap::new()),
         }
     }
 
@@ -232,6 +252,22 @@ impl TcpListenerManager {
             config.connection_rate_limit_per_minute,
         ));
         self.conn_semaphore = Arc::new(tokio::sync::Semaphore::new(config.max_connections as usize));
+
+        // Rebuild http_proxy with updated timeouts (shares the same ArcSwap<RouteManager>)
+        let mut http_proxy_svc = HttpProxyService::with_connect_timeout(
+            Arc::clone(&self.route_manager),
+            Arc::clone(&self.metrics),
+            std::time::Duration::from_millis(config.connection_timeout_ms),
+        );
+        http_proxy_svc.set_backend_tls_config(tls_handler::shared_backend_tls_config());
+        http_proxy_svc.set_backend_tls_config_alpn(tls_handler::shared_backend_tls_config_alpn());
+        http_proxy_svc.set_connection_timeouts(
+            std::time::Duration::from_millis(config.socket_timeout_ms),
+            std::time::Duration::from_millis(config.socket_timeout_ms),
+            std::time::Duration::from_millis(config.max_connection_lifetime_ms),
+        );
+        self.http_proxy = Arc::new(http_proxy_svc);
+
         self.conn_config = Arc::new(config);
     }
 
@@ -288,12 +324,13 @@ impl TcpListenerManager {
         let cancel = self.cancel_token.clone();
         let relay = Arc::clone(&self.socket_handler_relay);
         let semaphore = Arc::clone(&self.conn_semaphore);
+        let route_cancels = Arc::clone(&self.route_cancels);
 
         let handle = tokio::spawn(async move {
             Self::accept_loop(
                 listener, port, route_manager_swap, metrics, tls_configs,
                 shared_tls_acceptor, http_proxy, conn_config, conn_tracker, cancel, relay,
-                semaphore,
+                semaphore, route_cancels,
             ).await;
         });
 
@@ -336,13 +373,15 @@ impl TcpListenerManager {
 
         for (port, handle) in self.listeners.drain() {
             let remaining = deadline.saturating_duration_since(tokio::time::Instant::now());
+            let abort_handle = handle.abort_handle();
             if remaining.is_zero() {
-                handle.abort();
+                abort_handle.abort();
                 warn!("Force-stopped listener on port {} (timeout exceeded)", port);
             } else {
                 match tokio::time::timeout(remaining, handle).await {
                     Ok(_) => info!("Listener on port {} stopped gracefully", port),
                     Err(_) => {
+                        abort_handle.abort();
                         warn!("Listener on port {} did not stop in time, aborting", port);
                     }
                 }
@@ -370,6 +409,20 @@ impl TcpListenerManager {
         self.route_manager.store(route_manager);
     }
 
+    /// Cancel connections on routes that no longer exist in the active set.
+    /// Existing connections on removed routes are terminated via their per-route CancellationToken.
+    pub fn invalidate_removed_routes(&self, active_route_ids: &std::collections::HashSet<String>) {
+        self.route_cancels.retain(|id, token| {
+            if active_route_ids.contains(id) {
+                true
+            } else {
+                info!("Cancelling connections for removed route '{}'", id);
+                token.cancel();
+                false // remove cancelled token from map
+            }
+        });
+    }
+
     /// Prune HTTP proxy caches for route IDs that are no longer active.
     pub fn prune_http_proxy_caches(&self, active_route_ids: &std::collections::HashSet<String>) {
         self.http_proxy.prune_stale_routes(active_route_ids);
@@ -399,6 +452,7 @@ impl TcpListenerManager {
         cancel: CancellationToken,
         socket_handler_relay: Arc<std::sync::RwLock<Option<String>>>,
         conn_semaphore: Arc<tokio::sync::Semaphore>,
+        route_cancels: Arc<DashMap<String, CancellationToken>>,
     ) {
         loop {
             tokio::select! {
@@ -411,21 +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) => {
+                                    warn!("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) => {
+                                    warn!("Connection semaphore closed, dropping connection from {}", peer_addr);
                                     drop(stream);
                                     continue;
                                 }
@@ -433,7 +485,7 @@ impl TcpListenerManager {
 
                             // Check per-IP limits and rate limiting
                             if !conn_tracker.try_accept(&ip) {
-                                debug!("Rejected connection from {} (per-IP limit or rate limit)", peer_addr);
+                                warn!("Rejected connection from {} (per-IP limit or rate limit)", peer_addr);
                                 drop(stream);
                                 drop(permit);
                                 continue;
@@ -453,6 +505,7 @@ impl TcpListenerManager {
                             let ct = Arc::clone(&conn_tracker);
                             let cn = cancel.clone();
                             let sr = Arc::clone(&socket_handler_relay);
+                            let rc = Arc::clone(&route_cancels);
                             debug!("Accepted connection from {} on port {}", peer_addr, port);
 
                             tokio::spawn(async move {
@@ -461,10 +514,10 @@ impl TcpListenerManager {
                                 // RAII guard ensures connection_closed is called on all paths
                                 let _ct_guard = ConnectionTrackerGuard::new(ct, ip);
                                 let result = Self::handle_connection(
-                                    stream, port, peer_addr, rm, m, tc, sa, hp, cc, cn, sr,
+                                    stream, port, peer_addr, rm, m, tc, sa, hp, cc, cn, sr, rc,
                                 ).await;
                                 if let Err(e) = result {
-                                    debug!("Connection error from {}: {}", peer_addr, e);
+                                    warn!("Connection error from {}: {}", peer_addr, e);
                                 }
                             });
                         }
@@ -491,6 +544,7 @@ impl TcpListenerManager {
         conn_config: Arc<ConnectionConfig>,
         cancel: CancellationToken,
         socket_handler_relay: Arc<std::sync::RwLock<Option<String>>>,
+        route_cancels: Arc<DashMap<String, CancellationToken>>,
     ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
         use tokio::io::AsyncReadExt;
 
@@ -507,8 +561,9 @@ impl TcpListenerManager {
         // Non-proxy connections skip the peek entirely (no latency cost).
         let mut effective_peer_addr = peer_addr;
         if !conn_config.proxy_ips.is_empty() && conn_config.proxy_ips.contains(&peer_addr.ip()) {
-            // Trusted proxy IP — peek for PROXY protocol header
-            let mut proxy_peek = vec![0u8; 256];
+            // Trusted proxy IP — peek for PROXY protocol header.
+            // Use stack-allocated buffers (PROXY v1 headers are max ~108 bytes).
+            let mut proxy_peek = [0u8; 256];
             let pn = match tokio::time::timeout(
                 std::time::Duration::from_millis(conn_config.initial_data_timeout_ms),
                 stream.peek(&mut proxy_peek),
@@ -523,9 +578,9 @@ impl TcpListenerManager {
                     Ok((header, consumed)) => {
                         debug!("PROXY protocol: real client {} -> {}", header.source_addr, header.dest_addr);
                         effective_peer_addr = header.source_addr;
-                        // Consume the proxy protocol header bytes
-                        let mut discard = vec![0u8; consumed];
-                        stream.read_exact(&mut discard).await?;
+                        // Consume the proxy protocol header bytes (stack buffer, max 108 bytes)
+                        let mut discard = [0u8; 128];
+                        stream.read_exact(&mut discard[..consumed]).await?;
                     }
                     Err(e) => {
                         debug!("Failed to parse PROXY protocol header: {}", e);
@@ -595,12 +650,20 @@ impl TcpListenerManager {
                         let target_port = target.port.resolve(port);
                         let route_id = quick_match.route.id.as_deref();
 
+                        // Resolve per-route cancel token (child of global cancel)
+                        let conn_cancel = match route_id {
+                            Some(id) => route_cancels.entry(id.to_string())
+                                .or_insert_with(|| cancel.child_token())
+                                .clone(),
+                            None => cancel.clone(),
+                        };
+
                         // Check route-level IP security
                         if let Some(ref security) = quick_match.route.security {
                             if !rustproxy_http::request_filter::RequestFilter::check_ip_security(
                                 security, &peer_addr.ip(),
                             ) {
-                                debug!("Connection from {} blocked by route security", peer_addr);
+                                warn!("Connection from {} blocked by route security", peer_addr);
                                 return Ok(());
                             }
                         }
@@ -649,7 +712,7 @@ impl TcpListenerManager {
 
                             let (_bytes_in, _bytes_out) = forwarder::forward_bidirectional_with_timeouts(
                                 stream, backend_w, None,
-                                inactivity_timeout, max_lifetime, cancel,
+                                inactivity_timeout, max_lifetime, conn_cancel,
                                 Some(forwarder::ForwardMetricsCtx {
                                     collector: Arc::clone(&metrics),
                                     route_id: route_id.map(|s| s.to_string()),
@@ -659,7 +722,7 @@ impl TcpListenerManager {
                         } else {
                             let (_bytes_in, _bytes_out) = forwarder::forward_bidirectional_with_timeouts(
                                 stream, backend, None,
-                                inactivity_timeout, max_lifetime, cancel,
+                                inactivity_timeout, max_lifetime, conn_cancel,
                                 Some(forwarder::ForwardMetricsCtx {
                                     collector: Arc::clone(&metrics),
                                     route_id: route_id.map(|s| s.to_string()),
@@ -746,7 +809,7 @@ impl TcpListenerManager {
         let route_match = match route_match {
             Some(rm) => rm,
             None => {
-                debug!("No route matched for port {} domain {:?}", port, domain);
+                warn!("No route matched for port {} domain {:?} from {}", port, domain, peer_addr);
                 if is_http {
                     // Send a proper HTTP error instead of dropping the connection
                     use tokio::io::AsyncWriteExt;
@@ -764,13 +827,23 @@ impl TcpListenerManager {
 
         let route_id = route_match.route.id.as_deref();
 
+        // Resolve per-route cancel token (child of global cancel).
+        // When this route is removed via updateRoutes, the token is cancelled,
+        // terminating all connections on this route.
+        let cancel = match route_id {
+            Some(id) => route_cancels.entry(id.to_string())
+                .or_insert_with(|| cancel.child_token())
+                .clone(),
+            None => cancel,
+        };
+
         // Check route-level IP security for passthrough connections
         if let Some(ref security) = route_match.route.security {
             if !rustproxy_http::request_filter::RequestFilter::check_ip_security(
                 security,
                 &peer_addr.ip(),
             ) {
-                debug!("Connection from {} blocked by route security", peer_addr);
+                warn!("Connection from {} blocked by route security", peer_addr);
                 return Ok(());
             }
         }
@@ -791,7 +864,8 @@ impl TcpListenerManager {
                     stream, n, port, peer_addr,
                     &route_match, domain.as_deref(), is_tls,
                     &relay_socket_path,
-                    &metrics, route_id,
+                    Arc::clone(&metrics), route_id,
+                    &conn_config, cancel.clone(),
                 ).await;
             } else {
                 debug!("Socket-handler route matched but no relay path configured");
@@ -912,13 +986,18 @@ impl TcpListenerManager {
                     Err(_) => return Err("TLS handshake timeout".into()),
                 };
 
-                // Peek at decrypted data to determine if HTTP
+                // Peek at decrypted data to determine if HTTP.
+                // Timeout prevents connection leak if client completes TLS
+                // but never sends application data (scanners, health probes, slow-loris).
                 let mut buf_stream = tokio::io::BufReader::new(tls_stream);
                 let peeked = {
                     use tokio::io::AsyncBufReadExt;
-                    match buf_stream.fill_buf().await {
-                        Ok(data) => sni_parser::is_http(data),
-                        Err(_) => false,
+                    match tokio::time::timeout(
+                        std::time::Duration::from_millis(conn_config.initial_data_timeout_ms),
+                        buf_stream.fill_buf(),
+                    ).await {
+                        Ok(Ok(data)) => sni_parser::is_http(data),
+                        Ok(Err(_)) | Err(_) => false,
                     }
                 };
 
@@ -936,7 +1015,11 @@ impl TcpListenerManager {
                         "TLS Terminate + HTTP: {} -> {}:{} (domain: {:?})",
                         peer_addr, target_host, target_port, domain
                     );
-                    http_proxy.handle_io(buf_stream, peer_addr, port, cancel.clone()).await;
+                    // Wrap in ShutdownOnDrop to ensure TLS close_notify is sent
+                    // even if hyper drops the connection without calling shutdown
+                    // (e.g. H2 close, backend error, idle timeout drain).
+                    let wrapped = rustproxy_http::shutdown_on_drop::ShutdownOnDrop::new(buf_stream);
+                    http_proxy.handle_io(wrapped, peer_addr, port, cancel.clone()).await;
                 } else {
                     debug!(
                         "TLS Terminate + TCP: {} -> {}:{} (domain: {:?})",
@@ -964,7 +1047,7 @@ impl TcpListenerManager {
 
                     let (_bytes_in, _bytes_out) = Self::forward_bidirectional_split_with_timeouts(
                         tls_read, tls_write, backend_read, backend_write,
-                        inactivity_timeout, max_lifetime,
+                        inactivity_timeout, max_lifetime, cancel.clone(),
                         Some(forwarder::ForwardMetricsCtx {
                             collector: Arc::clone(&metrics),
                             route_id: route_id.map(|s| s.to_string()),
@@ -987,13 +1070,18 @@ impl TcpListenerManager {
                     Err(_) => return Err("TLS handshake timeout".into()),
                 };
 
-                // Peek at decrypted data to detect protocol
+                // Peek at decrypted data to detect protocol.
+                // Timeout prevents connection leak if client completes TLS
+                // but never sends application data (scanners, health probes, slow-loris).
                 let mut buf_stream = tokio::io::BufReader::new(tls_stream);
                 let is_http_data = {
                     use tokio::io::AsyncBufReadExt;
-                    match buf_stream.fill_buf().await {
-                        Ok(data) => sni_parser::is_http(data),
-                        Err(_) => false,
+                    match tokio::time::timeout(
+                        std::time::Duration::from_millis(conn_config.initial_data_timeout_ms),
+                        buf_stream.fill_buf(),
+                    ).await {
+                        Ok(Ok(data)) => sni_parser::is_http(data),
+                        Ok(Err(_)) | Err(_) => false,
                     }
                 };
 
@@ -1013,7 +1101,10 @@ impl TcpListenerManager {
                         "TLS Terminate+Reencrypt + HTTP: {} (domain: {:?})",
                         peer_addr, domain
                     );
-                    http_proxy.handle_io(buf_stream, peer_addr, port, cancel.clone()).await;
+                    // Wrap in ShutdownOnDrop to ensure TLS close_notify is sent
+                    // even if hyper drops the connection without calling shutdown.
+                    let wrapped = rustproxy_http::shutdown_on_drop::ShutdownOnDrop::new(buf_stream);
+                    http_proxy.handle_io(wrapped, peer_addr, port, cancel.clone()).await;
                 } else {
                     // Non-HTTP: TLS-to-TLS tunnel (existing behavior for raw TCP protocols)
                     debug!(
@@ -1023,7 +1114,7 @@ impl TcpListenerManager {
                     Self::handle_tls_reencrypt_tunnel(
                         buf_stream, &target_host, target_port,
                         peer_addr, Arc::clone(&metrics), route_id,
-                        &conn_config,
+                        &conn_config, cancel.clone(),
                     ).await?;
                 }
                 Ok(())
@@ -1100,8 +1191,10 @@ impl TcpListenerManager {
         domain: Option<&str>,
         is_tls: bool,
         relay_path: &str,
-        metrics: &MetricsCollector,
+        metrics: Arc<MetricsCollector>,
         route_id: Option<&str>,
+        conn_config: &ConnectionConfig,
+        cancel: CancellationToken,
     ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
         use tokio::io::{AsyncReadExt, AsyncWriteExt};
         use tokio::net::UnixStream;
@@ -1141,27 +1234,34 @@ impl TcpListenerManager {
         // Forward initial data to the Unix socket
         unix_stream.write_all(&initial_buf).await?;
 
-        // Bidirectional relay between TCP client and Unix socket handler
+        // Bidirectional relay with inactivity timeout, max lifetime, and cancellation.
+        // Split both streams and use the same watchdog pattern as other forwarding paths.
         let initial_len = initial_buf.len() as u64;
-        match tokio::io::copy_bidirectional(&mut stream, &mut unix_stream).await {
-            Ok((c2s, s2c)) => {
-                // Include initial data bytes that were forwarded before copy_bidirectional
-                let total_in = c2s + initial_len;
-                debug!("Socket handler relay complete for {}: {} bytes in, {} bytes out",
-                    route_key, total_in, s2c);
-                let ip = peer_addr.ip().to_string();
-                metrics.record_bytes(total_in, s2c, route_id, Some(&ip));
-            }
-            Err(e) => {
-                // Still record the initial data even on error
-                if initial_len > 0 {
-                    let ip = peer_addr.ip().to_string();
-                    metrics.record_bytes(initial_len, 0, route_id, Some(&ip));
-                }
-                debug!("Socket handler relay ended for {}: {}", route_key, e);
-            }
+        let inactivity_timeout = std::time::Duration::from_millis(conn_config.socket_timeout_ms);
+        let max_lifetime = std::time::Duration::from_millis(conn_config.max_connection_lifetime_ms);
+
+        let (tcp_read, tcp_write) = stream.into_split();
+        let (unix_read, unix_write) = unix_stream.into_split();
+
+        let ip_str = peer_addr.ip().to_string();
+        let (_bytes_in, _bytes_out) = Self::forward_bidirectional_split_with_timeouts(
+            tcp_read, tcp_write, unix_read, unix_write,
+            inactivity_timeout, max_lifetime, cancel,
+            Some(forwarder::ForwardMetricsCtx {
+                collector: Arc::clone(&metrics),
+                route_id: route_id.map(|s| s.to_string()),
+                source_ip: Some(ip_str.clone()),
+            }),
+        ).await;
+
+        // Include the initial data that was forwarded before the bidirectional relay
+        if initial_len > 0 {
+            metrics.record_bytes(initial_len, 0, route_id, Some(&ip_str));
         }
 
+        debug!("Socket handler relay complete for {}: {} bytes in, {} bytes out",
+            route_key, _bytes_in + initial_len, _bytes_out);
+
         Ok(())
     }
 
@@ -1176,6 +1276,7 @@ impl TcpListenerManager {
         metrics: Arc<MetricsCollector>,
         route_id: Option<&str>,
         conn_config: &ConnectionConfig,
+        cancel: CancellationToken,
     ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
         // Connect to backend over TLS with timeout
         let backend_tls = match tokio::time::timeout(
@@ -1220,7 +1321,7 @@ impl TcpListenerManager {
 
         let (_bytes_in, _bytes_out) = Self::forward_bidirectional_split_with_timeouts(
             client_read, client_write, backend_read, backend_write,
-            inactivity_timeout, max_lifetime,
+            inactivity_timeout, max_lifetime, cancel,
             Some(forwarder::ForwardMetricsCtx {
                 collector: metrics,
                 route_id: route_id.map(|s| s.to_string()),
@@ -1295,6 +1396,7 @@ impl TcpListenerManager {
         mut backend_write: W2,
         inactivity_timeout: std::time::Duration,
         max_lifetime: std::time::Duration,
+        cancel: CancellationToken,
         metrics: Option<forwarder::ForwardMetricsCtx>,
     ) -> (u64, u64)
     where
@@ -1310,15 +1412,24 @@ 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;
@@ -1332,19 +1443,27 @@ 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;
@@ -1358,42 +1477,57 @@ 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 and max lifetime
+        // 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;
             loop {
-                tokio::time::sleep(check_interval).await;
-
-                // Check max lifetime
-                if start.elapsed() >= max_lifetime {
-                    debug!("Connection exceeded max lifetime, closing");
-                    c2b_handle.abort();
-                    b2c_handle.abort();
-                    break;
-                }
-
-                // Check inactivity
-                let current = la_watch.load(Ordering::Relaxed);
-                if current == last_seen {
-                    // No activity since last check
-                    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();
+                tokio::select! {
+                    _ = cancel.cancelled() => {
+                        debug!("Split-stream connection cancelled by shutdown");
                         break;
                     }
+                    _ = tokio::time::sleep(check_interval) => {
+                        // Check max lifetime
+                        if start.elapsed() >= max_lifetime {
+                            debug!("Connection exceeded max lifetime, closing");
+                            break;
+                        }
+
+                        // Check inactivity
+                        let current = la_watch.load(Ordering::Relaxed);
+                        if current == last_seen {
+                            // No activity since last check
+                            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);
+                                break;
+                            }
+                        }
+                        last_seen = current;
+                    }
                 }
-                last_seen = current;
             }
+            // 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);

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

@@ -98,10 +98,24 @@ pub fn build_shared_tls_acceptor(resolver: CertResolver) -> Result<TlsAcceptor,
 }
 
 /// Build a TLS acceptor from PEM-encoded cert and key data.
+/// Advertises both h2 and http/1.1 via ALPN (for client-facing connections).
 pub fn build_tls_acceptor(cert_pem: &str, key_pem: &str) -> Result<TlsAcceptor, Box<dyn std::error::Error + Send + Sync>> {
     build_tls_acceptor_with_config(cert_pem, key_pem, None)
 }
 
+/// Build a TLS acceptor for backend servers that only speak HTTP/1.1.
+/// Does NOT advertise h2 in ALPN, preventing false h2 auto-detection.
+pub fn build_tls_acceptor_h1_only(cert_pem: &str, key_pem: &str) -> Result<TlsAcceptor, Box<dyn std::error::Error + Send + Sync>> {
+    ensure_crypto_provider();
+    let certs = load_certs(cert_pem)?;
+    let key = load_private_key(key_pem)?;
+    let mut config = ServerConfig::builder()
+        .with_no_client_auth()
+        .with_single_cert(certs, key)?;
+    config.alpn_protocols = vec![b"http/1.1".to_vec()];
+    Ok(TlsAcceptor::from(Arc::new(config)))
+}
+
 /// Build a TLS acceptor with optional RouteTls configuration for version/cipher tuning.
 pub fn build_tls_acceptor_with_config(
     cert_pem: &str,
@@ -204,6 +218,25 @@ pub fn shared_backend_tls_config() -> Arc<rustls::ClientConfig> {
     }).clone()
 }
 
+/// Get or create a shared backend TLS `ClientConfig` with ALPN `h2` + `http/1.1`.
+///
+/// Used for auto-detection mode: the backend server picks its preferred protocol
+/// via ALPN, and the proxy reads the negotiated result to decide h1 vs h2 forwarding.
+static SHARED_CLIENT_CONFIG_ALPN: OnceLock<Arc<rustls::ClientConfig>> = OnceLock::new();
+
+pub fn shared_backend_tls_config_alpn() -> Arc<rustls::ClientConfig> {
+    SHARED_CLIENT_CONFIG_ALPN.get_or_init(|| {
+        ensure_crypto_provider();
+        let mut config = rustls::ClientConfig::builder()
+            .dangerous()
+            .with_custom_certificate_verifier(Arc::new(InsecureVerifier))
+            .with_no_client_auth();
+        config.alpn_protocols = vec![b"h2".to_vec(), b"http/1.1".to_vec()];
+        info!("Built shared backend TLS client config with ALPN h2+http/1.1 for auto-detection");
+        Arc::new(config)
+    }).clone()
+}
+
 /// Connect to a backend with TLS (for terminate-and-reencrypt mode).
 /// Uses the shared backend TLS config for session resumption.
 pub async fn connect_tls(

rust/crates/rustproxy-routing/src/matchers/domain.rs

@@ -6,25 +6,28 @@
 /// - `example.com` exact match
 /// - `**.example.com` matches any depth of subdomain
 pub fn domain_matches(pattern: &str, domain: &str) -> bool {
-    let pattern = pattern.trim().to_lowercase();
-    let domain = domain.trim().to_lowercase();
+    let pattern = pattern.trim();
+    let domain = domain.trim();
 
     if pattern == "*" {
         return true;
     }
 
-    if pattern == domain {
+    if pattern.eq_ignore_ascii_case(domain) {
         return true;
     }
 
     // Wildcard patterns
-    if pattern.starts_with("*.") {
+    if pattern.starts_with("*.") || pattern.starts_with("*.") {
         let suffix = &pattern[2..]; // e.g., "example.com"
         // Match exact parent or any single-level subdomain
-        if domain == suffix {
+        if domain.eq_ignore_ascii_case(suffix) {
             return true;
         }
-        if domain.ends_with(&format!(".{}", suffix)) {
+        if domain.len() > suffix.len() + 1
+            && domain.as_bytes()[domain.len() - suffix.len() - 1] == b'.'
+            && domain[domain.len() - suffix.len()..].eq_ignore_ascii_case(suffix)
+        {
             // Check it's a single level subdomain for `*.`
             let prefix = &domain[..domain.len() - suffix.len() - 1];
             return !prefix.contains('.');
@@ -35,11 +38,22 @@ pub fn domain_matches(pattern: &str, domain: &str) -> bool {
     if pattern.starts_with("**.") {
         let suffix = &pattern[3..];
         // Match exact parent or any depth of subdomain
-        return domain == suffix || domain.ends_with(&format!(".{}", suffix));
+        if domain.eq_ignore_ascii_case(suffix) {
+            return true;
+        }
+        if domain.len() > suffix.len() + 1
+            && domain.as_bytes()[domain.len() - suffix.len() - 1] == b'.'
+            && domain[domain.len() - suffix.len()..].eq_ignore_ascii_case(suffix)
+        {
+            return true;
+        }
+        return false;
     }
 
-    // Use glob-match for more complex patterns
-    glob_match::glob_match(&pattern, &domain)
+    // Use glob-match for more complex patterns (case-insensitive via lowercasing)
+    let pattern_lower = pattern.to_lowercase();
+    let domain_lower = domain.to_lowercase();
+    glob_match::glob_match(&pattern_lower, &domain_lower)
 }
 
 /// Check if a domain matches any of the given patterns.

rust/crates/rustproxy-routing/src/route_manager.rs

@@ -60,6 +60,16 @@ impl RouteManager {
         manager
     }
 
+    /// Check if any route on the given port uses header matching.
+    /// Used to skip expensive header HashMap construction when no route needs it.
+    pub fn any_route_has_headers(&self, port: u16) -> bool {
+        if let Some(indices) = self.port_index.get(&port) {
+            indices.iter().any(|&idx| self.routes[idx].route_match.headers.is_some())
+        } else {
+            false
+        }
+    }
+
     /// Find the best matching route for the given context.
     pub fn find_route<'a>(&'a self, ctx: &MatchContext<'_>) -> Option<RouteMatchResult<'a>> {
         // Get routes for this port

rust/crates/rustproxy/src/lib.rs

@@ -603,17 +603,42 @@ impl RustProxy {
             .collect();
         self.metrics.retain_routes(&active_route_ids);
 
+        // Prune per-backend metrics for backends no longer in any route target.
+        // For PortSpec::Preserve routes, expand across all listening ports since
+        // the actual runtime port depends on the incoming connection.
+        let listening_ports = self.get_listening_ports();
+        let active_backends: HashSet<String> = routes.iter()
+            .filter_map(|r| r.action.targets.as_ref())
+            .flat_map(|targets| targets.iter())
+            .flat_map(|target| {
+                let hosts: Vec<String> = target.host.to_vec().into_iter().map(|s| s.to_string()).collect();
+                match &target.port {
+                    rustproxy_config::PortSpec::Fixed(p) => {
+                        hosts.into_iter().map(|h| format!("{}:{}", h, p)).collect::<Vec<_>>()
+                    }
+                    _ => {
+                        // Preserve/special: expand across all listening ports
+                        let lp = &listening_ports;
+                        hosts.into_iter()
+                            .flat_map(|h| lp.iter().map(move |p| format!("{}:{}", h, *p)))
+                            .collect::<Vec<_>>()
+                    }
+                }
+            })
+            .collect();
+        self.metrics.retain_backends(&active_backends);
+
         // Atomically swap the route table
         let new_manager = Arc::new(new_manager);
         self.route_table.store(Arc::clone(&new_manager));
 
-        // Update listener manager
+        // Update listener manager.
+        // IMPORTANT: TLS configs must be swapped BEFORE the route manager so that
+        // new routes only become visible after their certs are loaded. The reverse
+        // order (routes first) creates a window where connections match new routes
+        // but get the old TLS acceptor, causing cert mismatches.
         if let Some(ref mut listener) = self.listener_manager {
-            listener.update_route_manager(Arc::clone(&new_manager));
-            // Prune HTTP proxy caches (rate limiters, regex cache, round-robin counters)
-            listener.prune_http_proxy_caches(&active_route_ids);
-
-            // Update TLS configs
+            // 1. Update TLS configs first (so new certs are available before new routes)
             let mut tls_configs = Self::extract_tls_configs(&routes);
             if let Some(ref cm_arc) = self.cert_manager {
                 let cm = cm_arc.lock().await;
@@ -634,6 +659,13 @@ impl RustProxy {
             }
             listener.set_tls_configs(tls_configs);
 
+            // 2. Now swap the route manager (new routes become visible with certs already loaded)
+            listener.update_route_manager(Arc::clone(&new_manager));
+            // Cancel connections on routes that were removed or disabled
+            listener.invalidate_removed_routes(&active_route_ids);
+            // Prune HTTP proxy caches (rate limiters, regex cache, round-robin counters)
+            listener.prune_http_proxy_caches(&active_route_ids);
+
             // Add new ports
             for port in &new_ports {
                 if !old_ports.contains(port) {

rust/crates/rustproxy/tests/common/mod.rs

@@ -195,7 +195,10 @@ pub async fn start_tls_http_backend(
 ) -> JoinHandle<()> {
     use std::sync::Arc;
 
-    let acceptor = rustproxy_passthrough::build_tls_acceptor(cert_pem, key_pem)
+    // Use h1-only acceptor: test backends speak raw HTTP/1.1 text,
+    // so they must NOT advertise h2 via ALPN (which would cause
+    // auto-detect to attempt h2 binary framing and fail).
+    let acceptor = rustproxy_passthrough::build_tls_acceptor_h1_only(cert_pem, key_pem)
         .expect("Failed to build TLS acceptor");
     let acceptor = Arc::new(acceptor);
     let name = backend_name.to_string();

test/test.sni-requirement.node.ts

@@ -7,10 +7,15 @@
 import { expect, tap } from '@git.zone/tstest/tapbundle';
 import { SmartProxy } from '../ts/proxies/smart-proxy/index.js';
 import type { IRouteConfig } from '../ts/proxies/smart-proxy/models/route-types.js';
+import { findFreePorts } from './helpers/port-allocator.js';
 
-// Use unique high ports for each test to avoid conflicts
-let testPort = 20000;
-const getNextPort = () => testPort++;
+let testPorts: number[];
+let portIndex = 0;
+const getNextPort = () => testPorts[portIndex++];
+
+tap.test('setup - allocate ports', async () => {
+  testPorts = await findFreePorts(16);
+});
 
 // --------------------------------- Single Route, No Domain Restriction ---------------------------------
 

ts/00_commitinfo_data.ts

@@ -3,6 +3,6 @@
  */
 export const commitinfo = {
   name: '@push.rocks/smartproxy',
-  version: '25.8.3',
+  version: '25.11.14',
   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.'
 }

ts/proxies/smart-proxy/models/interfaces.ts

@@ -112,12 +112,12 @@ export interface ISmartProxyOptions {
   maxVersion?: string;
 
   // Timeout settings
-  connectionTimeout?: number; // Timeout for establishing connection to backend (ms), default: 30000 (30s)
+  connectionTimeout?: number; // Timeout for establishing connection to backend (ms), default: 60000 (60s)
   initialDataTimeout?: number; // Timeout for initial data/SNI (ms), default: 60000 (60s)
-  socketTimeout?: number; // Socket inactivity timeout (ms), default: 3600000 (1h)
+  socketTimeout?: number; // Socket inactivity timeout (ms), default: 60000 (60s)
   inactivityCheckInterval?: number; // How often to check for inactive connections (ms), default: 60000 (60s)
-  maxConnectionLifetime?: number; // Default max connection lifetime (ms), default: 86400000 (24h)
-  inactivityTimeout?: number; // Inactivity timeout (ms), default: 14400000 (4h)
+  maxConnectionLifetime?: number; // Max connection lifetime (ms), default: 3600000 (1h)
+  inactivityTimeout?: number; // Inactivity timeout (ms), default: 75000 (75s)
 
   gracefulShutdownTimeout?: number; // (ms) maximum time to wait for connections to close during shutdown
 

ts/proxies/smart-proxy/models/metrics-types.ts

@@ -67,6 +67,13 @@ export interface IMetrics {
     connections(): number;
   };
   
+  // Backend metrics
+  backends: {
+    byBackend(): Map<string, IBackendMetrics>;
+    protocols(): Map<string, string>;
+    topByErrors(limit?: number): Array<{ backend: string; errors: number }>;
+  };
+
   // Performance metrics
   percentiles: {
     connectionDuration(): { p50: number; p95: number; p99: number };
@@ -98,6 +105,21 @@ export interface IMetricsConfig {
   prometheusPrefix: string;        // Default: smartproxy_
 }
 
+/**
+ * Per-backend metrics
+ */
+export interface IBackendMetrics {
+  protocol: string;
+  activeConnections: number;
+  totalConnections: number;
+  connectErrors: number;
+  handshakeErrors: number;
+  requestErrors: number;
+  avgConnectTimeMs: number;
+  poolHitRate: number;
+  h2Failures: number;
+}
+
 /**
  * Internal interface for connection byte tracking
  */

ts/proxies/smart-proxy/models/route-types.ts

@@ -262,7 +262,7 @@ export interface IRouteAction {
   
   // Additional options for backend-specific settings
   options?: {
-    backendProtocol?: 'http1' | 'http2';
+    backendProtocol?: 'http1' | 'http2' | 'auto';
     [key: string]: any;
   };
 

ts/proxies/smart-proxy/rust-metrics-adapter.ts

@@ -1,4 +1,4 @@
-import type { IMetrics, IThroughputData, IThroughputHistoryPoint } from './models/metrics-types.js';
+import type { IMetrics, IBackendMetrics, IThroughputData, IThroughputHistoryPoint } from './models/metrics-types.js';
 import type { RustProxyBridge } from './rust-proxy-bridge.js';
 
 /**
@@ -169,6 +169,55 @@ export class RustMetricsAdapter implements IMetrics {
     },
   };
 
+  public backends = {
+    byBackend: (): Map<string, IBackendMetrics> => {
+      const result = new Map<string, IBackendMetrics>();
+      if (this.cache?.backends) {
+        for (const [key, bm] of Object.entries(this.cache.backends)) {
+          const m = bm as any;
+          const totalTimeUs = m.totalConnectTimeUs ?? 0;
+          const count = m.connectCount ?? 0;
+          const poolHits = m.poolHits ?? 0;
+          const poolMisses = m.poolMisses ?? 0;
+          const poolTotal = poolHits + poolMisses;
+          result.set(key, {
+            protocol: m.protocol ?? 'unknown',
+            activeConnections: m.activeConnections ?? 0,
+            totalConnections: m.totalConnections ?? 0,
+            connectErrors: m.connectErrors ?? 0,
+            handshakeErrors: m.handshakeErrors ?? 0,
+            requestErrors: m.requestErrors ?? 0,
+            avgConnectTimeMs: count > 0 ? (totalTimeUs / count) / 1000 : 0,
+            poolHitRate: poolTotal > 0 ? poolHits / poolTotal : 0,
+            h2Failures: m.h2Failures ?? 0,
+          });
+        }
+      }
+      return result;
+    },
+    protocols: (): Map<string, string> => {
+      const result = new Map<string, string>();
+      if (this.cache?.backends) {
+        for (const [key, bm] of Object.entries(this.cache.backends)) {
+          result.set(key, (bm as any).protocol ?? 'unknown');
+        }
+      }
+      return result;
+    },
+    topByErrors: (limit: number = 10): Array<{ backend: string; errors: number }> => {
+      const result: Array<{ backend: string; errors: number }> = [];
+      if (this.cache?.backends) {
+        for (const [key, bm] of Object.entries(this.cache.backends)) {
+          const m = bm as any;
+          const errors = (m.connectErrors ?? 0) + (m.handshakeErrors ?? 0) + (m.requestErrors ?? 0);
+          if (errors > 0) result.push({ backend: key, errors });
+        }
+      }
+      result.sort((a, b) => b.errors - a.errors);
+      return result.slice(0, limit);
+    },
+  };
+
   public percentiles = {
     connectionDuration: (): { p50: number; p95: number; p99: number } => {
       return { p50: 0, p95: 0, p99: 0 };

ts/proxies/smart-proxy/smart-proxy.ts

@@ -47,16 +47,16 @@ export class SmartProxy extends plugins.EventEmitter {
     // Apply defaults
     this.settings = {
       ...settingsArg,
-      initialDataTimeout: settingsArg.initialDataTimeout || 120000,
-      socketTimeout: settingsArg.socketTimeout || 3600000,
-      maxConnectionLifetime: settingsArg.maxConnectionLifetime || 86400000,
-      inactivityTimeout: settingsArg.inactivityTimeout || 14400000,
-      gracefulShutdownTimeout: settingsArg.gracefulShutdownTimeout || 30000,
+      initialDataTimeout: settingsArg.initialDataTimeout || 60_000,
+      socketTimeout: settingsArg.socketTimeout || 60_000,
+      maxConnectionLifetime: settingsArg.maxConnectionLifetime || 3_600_000,
+      inactivityTimeout: settingsArg.inactivityTimeout || 75_000,
+      gracefulShutdownTimeout: settingsArg.gracefulShutdownTimeout || 30_000,
       maxConnectionsPerIP: settingsArg.maxConnectionsPerIP || 100,
       connectionRateLimitPerMinute: settingsArg.connectionRateLimitPerMinute || 300,
-      keepAliveTreatment: settingsArg.keepAliveTreatment || 'extended',
-      keepAliveInactivityMultiplier: settingsArg.keepAliveInactivityMultiplier || 6,
-      extendedKeepAliveLifetime: settingsArg.extendedKeepAliveLifetime || 7 * 24 * 60 * 60 * 1000,
+      keepAliveTreatment: settingsArg.keepAliveTreatment || 'standard',
+      keepAliveInactivityMultiplier: settingsArg.keepAliveInactivityMultiplier || 4,
+      extendedKeepAliveLifetime: settingsArg.extendedKeepAliveLifetime || 3_600_000,
     };
 
     // Normalize ACME options

ts/proxies/smart-proxy/socket-handler-server.ts

@@ -92,6 +92,16 @@ export class SocketHandlerServer {
     let metadataBuffer = '';
     let metadataParsed = false;
 
+    // 10s timeout for metadata parsing phase — if Rust connects but never
+    // sends the JSON metadata line, don't hold the socket open indefinitely.
+    socket.setTimeout(10_000);
+    socket.on('timeout', () => {
+      if (!metadataParsed) {
+        logger.log('warn', 'Socket handler metadata timeout, closing', { component: 'socket-handler-server' });
+        socket.destroy();
+      }
+    });
+
     const onData = (chunk: Buffer) => {
       if (metadataParsed) return;
 
@@ -108,6 +118,7 @@ export class SocketHandlerServer {
       }
 
       metadataParsed = true;
+      socket.setTimeout(0); // Clear metadata timeout
       socket.removeListener('data', onData);
       socket.pause(); // Prevent data loss between handler removal and pipe setup