v25.11.19

changelog.md

@@ -1,5 +1,12 @@
 # Changelog
 
+## 2026-03-16 - 25.11.19 - fix(rustproxy-http)
+avoid reusing pooled HTTP/2 connections for requests with bodies to prevent upload flow-control stalls
+
+- Limit HTTP/2 pool checkout to bodyless requests such as GET, HEAD, and DELETE
+- Skip re-registering HTTP/2 connections in the pool after requests that send a body
+- Prevent stalled uploads caused by depleted connection-level flow control windows on reused HTTP/2 connections
+
 ## 2026-03-16 - 25.11.18 - fix(repo)
 no changes to commit
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@push.rocks/smartproxy",
-  "version": "25.11.18",
+  "version": "25.11.19",
   "private": false,
   "description": "A powerful proxy package with unified route-based configuration for high traffic management. Features include SSL/TLS support, flexible routing patterns, WebSocket handling, advanced security options, and automatic ACME certificate management.",
   "main": "dist_ts/index.js",

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

@@ -677,12 +677,20 @@ impl HttpProxyService {
                 h2: use_h2,
             };
 
-            // H2 pool checkout with async readiness validation.
-            // checkout_h2 does synchronous is_closed()/is_ready() checks, but these
-            // reflect cached state — the H2 connection driver (a separate tokio task)
-            // may not have processed a pending GOAWAY/RST yet. The ready().await
-            // forces the runtime to yield, giving the driver a chance to detect failures.
-            if use_h2 {
+            // H2 pool checkout — only for bodyless requests (GET/HEAD/DELETE).
+            //
+            // WORKAROUND: Requests with bodies (POST/PUT uploads) always get fresh H2
+            // connections. Reusing a pooled H2 connection after a large upload can stall
+            // forever due to depleted connection-level flow control windows. The h2 crate
+            // has no stall/timeout detection (https://github.com/hyperium/hyper/issues/2899),
+            // and Go/nginx HTTP/2 servers have known issues with connection-level window
+            // replenishment after large transfers (https://github.com/golang/go/issues/16481,
+            // https://github.com/golang/go/issues/56558). A fresh connection guarantees
+            // clean flow control state. The overhead is ~3-5ms for TLS+H2 handshake.
+            //
+            // TODO: Revisit once h2 crate adds flow control stall detection, or once
+            // Go/nginx H2 connection-level window handling is confirmed reliable.
+            if use_h2 && body.is_end_stream() {
                 if let Some((mut sender, age)) = self.connection_pool.checkout_h2(&pool_key) {
                     match tokio::time::timeout(
                         std::time::Duration::from_millis(500),
@@ -1040,10 +1048,12 @@ impl HttpProxyService {
             });
         }
 
-        // Clone sender for potential pool registration; register only after first request succeeds
+        // Only pool the H2 connection if the request had no body.
+        // Requests with bodies (uploads) deplete connection-level flow control windows.
+        let request_had_body = !body.is_end_stream();
         let sender_for_pool = sender.clone();
         let result = self.forward_h2_with_sender(sender, parts, body, upstream_headers, upstream_path, route, route_id, source_ip, Some(pool_key), domain, conn_activity).await;
-        if matches!(&result, Ok(ref resp) if resp.status() != StatusCode::BAD_GATEWAY) {
+        if !request_had_body && matches!(&result, Ok(ref resp) if resp.status() != StatusCode::BAD_GATEWAY) {
             let g = self.connection_pool.register_h2(pool_key.clone(), sender_for_pool);
             gen_holder.store(g, std::sync::atomic::Ordering::Relaxed);
         }
@@ -1369,9 +1379,12 @@ impl HttpProxyService {
 
                 match sender.send_request(upstream_req).await {
                     Ok(upstream_response) => {
-                        // H2 works! Register sender in pool for multiplexed reuse
-                        let g = self.connection_pool.register_h2(pool_key.clone(), sender);
-                        gen_holder.store(g, std::sync::atomic::Ordering::Relaxed);
+                        // Only pool after bodyless requests — uploads deplete connection-level
+                        // flow control windows (see comment at pool checkout above).
+                        if retry_state.is_some() {
+                            let g = self.connection_pool.register_h2(pool_key.clone(), sender);
+                            gen_holder.store(g, std::sync::atomic::Ordering::Relaxed);
+                        }
                         self.build_streaming_response(upstream_response, route, route_id, source_ip, conn_activity).await
                     }
                     Err(e) => {

ts/00_commitinfo_data.ts

@@ -3,6 +3,6 @@
  */
 export const commitinfo = {
   name: '@push.rocks/smartproxy',
-  version: '25.11.18',
+  version: '25.11.19',
   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.'
 }