fix(proxy-tunnel): 修复隧道连接池竞态与跨平台兼容问题

- 修复 TCP keepalive with_retries 在 Windows 上不可用的编译问题
- dispatcher 中 try_send 失败时记录警告日志而非静默丢弃
- 优化 handler_handles 清理策略，每 64 帧定期清理
- StreamState 记住原始连接引用，清理时避免连接池竞态

aether-proxy/src/tunnel/client.rs

@@ -170,8 +170,9 @@ fn configure_tcp_socket(stream: &TcpStream, state: &Arc<AppState>) {
     if state.config.tunnel_tcp_keepalive_secs > 0 {
         let keepalive = socket2::TcpKeepalive::new()
             .with_time(Duration::from_secs(state.config.tunnel_tcp_keepalive_secs))
-            .with_interval(Duration::from_secs(5))
+            .with_interval(Duration::from_secs(5));
-            .with_retries(3);
+        #[cfg(not(target_os = "windows"))]
+        let keepalive = keepalive.with_retries(3);
         if let Err(e) = sock_ref.set_tcp_keepalive(&keepalive) {
             warn!(error = %e, "failed to set TCP keepalive on tunnel socket");
         }

aether-proxy/src/tunnel/dispatcher.rs

@@ -37,6 +37,7 @@ where
     // Track spawned stream handlers so we can wait for them on shutdown
     let mut handler_handles: Vec<JoinHandle<()>> = Vec::new();
     let max_streams = state.config.tunnel_max_streams.unwrap_or(128) as usize;
+    let mut frames_since_cleanup: u32 = 0;
     let stale_timeout = Duration::from_secs(state.config.tunnel_stale_timeout_secs);
 
     // Track last time we received any data to detect stale connections
@@ -97,12 +98,20 @@ where
                     Err(e) => {
                         warn!(stream_id = frame.stream_id, error = %e, "invalid request metadata");
                         // Use try_send to avoid blocking the read loop
-                        let _ = frame_tx.try_send(Frame::new(
+                        if frame_tx
-                            frame.stream_id,
+                            .try_send(Frame::new(
-                            MsgType::StreamError,
+                                frame.stream_id,
-                            0,
+                                MsgType::StreamError,
-                            Bytes::from(format!("invalid request metadata: {e}")),
+                                0,
-                        ));
+                                Bytes::from(format!("invalid request metadata: {e}")),
+                            ))
+                            .is_err()
+                        {
+                            warn!(
+                                stream_id = frame.stream_id,
+                                "writer channel full, StreamError dropped"
+                            );
+                        }
                         continue;
                     }
                 };
@@ -112,12 +121,20 @@ where
                         stream_id = frame.stream_id,
                         "max concurrent streams reached"
                     );
-                    let _ = frame_tx.try_send(Frame::new(
+                    if frame_tx
-                        frame.stream_id,
+                        .try_send(Frame::new(
-                        MsgType::StreamError,
+                            frame.stream_id,
-                        0,
+                            MsgType::StreamError,
-                        Bytes::from("max concurrent streams reached"),
+                            0,
-                    ));
+                            Bytes::from("max concurrent streams reached"),
+                        ))
+                        .is_err()
+                    {
+                        warn!(
+                            stream_id = frame.stream_id,
+                            "writer channel full, StreamError dropped"
+                        );
+                    }
                     continue;
                 }
 
@@ -163,7 +180,12 @@ where
 
             MsgType::Ping => {
                 // Use try_send to avoid blocking the read loop when writer is congested
-                let _ = frame_tx.try_send(Frame::control(MsgType::Pong, frame.payload));
+                if frame_tx
+                    .try_send(Frame::control(MsgType::Pong, frame.payload))
+                    .is_err()
+                {
+                    warn!("writer channel full, Pong dropped");
+                }
             }
 
             MsgType::HeartbeatAck => {
@@ -180,9 +202,12 @@ where
             }
         }
 
-        // Periodically clean up finished handles to avoid unbounded growth
+        // Periodically clean up finished handles to avoid unbounded growth.
-        if handler_handles.len() > max_streams {
+        // Trigger every 64 frames OR when the count exceeds max_streams.
+        frames_since_cleanup += 1;
+        if frames_since_cleanup >= 64 || handler_handles.len() > max_streams {
             handler_handles.retain(|h| !h.is_finished());
+            frames_since_cleanup = 0;
         }
     };
 

src/services/proxy_node/tunnel_manager.py

@@ -63,7 +63,7 @@ class TunnelConnection:
             raise TunnelStreamError("frame send timeout (writer congested)")
 
     def create_stream(self, stream_id: int) -> _StreamState:
-        state = _StreamState(stream_id)
+        state = _StreamState(stream_id, conn=self)
         self._pending_streams[stream_id] = state
         return state
 
@@ -109,9 +109,10 @@ class _StreamState:
         "_body_chunks",
         "_done_event",
         "_error",
+        "_conn",
     )
 
-    def __init__(self, stream_id: int) -> None:
+    def __init__(self, stream_id: int, conn: TunnelConnection | None = None) -> None:
         self.stream_id = stream_id
         self.status: int = 0
         self.headers: list[list[str]] = []
@@ -119,6 +120,7 @@ class _StreamState:
         self._body_chunks: asyncio.Queue[bytes | None] = asyncio.Queue()
         self._done_event = asyncio.Event()
         self._error: str | None = None
+        self._conn = conn
 
     def set_response_headers(self, status: int, headers: list[list[str]] | dict[str, str]) -> None:
         self.status = status

src/services/proxy_node/tunnel_transport.py

@@ -95,7 +95,10 @@ class TunnelTransport(httpx.AsyncBaseTransport):
     def _cleanup_stream(self, manager: TunnelManager, stream_state: _StreamState | None) -> None:
         if stream_state is None:
             return
-        conn = manager.get_connection(self._node_id)
+        # 优先从 stream 记住的原始连接上移除，避免连接池竞态
+        conn = stream_state._conn
+        if conn is None:
+            conn = manager.get_connection(self._node_id)
         if conn:
             conn.remove_stream(stream_state.stream_id)
 
@@ -120,8 +123,10 @@ class TunnelResponseStream(httpx.AsyncByteStream):
             yield chunk
 
     async def aclose(self) -> None:
-        # 确保 stream 从 connection 的 pending 列表中移除，防止内存泄漏
+        # 从 stream 记住的原始连接上精确移除，避免连接池竞态
-        conn = self._manager.get_connection(self._node_id)
+        conn = self._stream_state._conn
+        if conn is None:
+            conn = self._manager.get_connection(self._node_id)
         if conn:
             conn.remove_stream(self._stream_state.stream_id)