feat: support native streaming responses

CHANGELOG.md

@@ -6,6 +6,7 @@
 - 新增独立进程并发与长连接内存测量工具，记录成功吞吐、P95/P99、服务端分配、空闲连接堆内存和回收情况；修正并行 `ns/op` 与客户端/服务端混合分配的解释。
 - 加固 native HTTP/1.1 解析边界：校验 Host、重复 Content-Length、Connection token、Transfer-Encoding、控制字符和响应头名称；补充 HTTP/1.0 Keep-Alive、HEAD、204 以及异常请求测试。
 - native 支持带 trailers 的 chunked 请求体，并继续通过 `MaxBodyBytes`、读超时和头部上限控制解析内存与连接占用。
+- native 响应实现 `http.Flusher`：HTTP/1.1 流式响应自动使用 chunked framing，普通响应继续自动生成 `Content-Length`。
 - 补充慢请求体超时、大请求体和同连接管线请求验收，确认请求体未完整读取时不会进入业务 Handler，连续请求不会串线。
 - 补充 `MaxConnections` 满载拒绝与槽位释放验收，确认请求完成后连接容量可再次使用。
 - 增加 8 并发、800 次建连/断连的状态池 churn 验收，覆盖连接对象回收与容量复用。

README.md

@@ -121,7 +121,7 @@ if err := app.RunWith(ctx, native.New(nil), ":8080"); err != nil {
 }
 ```
 
-`transport/native` 当前支持 HTTP/1.1、Keep-Alive、`Content-Length` 和 chunked 请求体、读写超时和优雅关闭；不支持的传输编码会被拒绝。它会复用连接状态和响应对象，并可通过 `MaxConnections`、`MaxConcurrentRequests` 限制内存峰值。它仍通过标准 HTTP 请求/响应接口连接业务层，适合在现有 Handler / Service 上做网络层验证。
+`transport/native` 当前支持 HTTP/1.1、Keep-Alive、`Content-Length` 和 chunked 请求体、`http.Flusher` 流式响应、读写超时和优雅关闭；不支持的传输编码会被拒绝。它会复用连接状态和响应对象，并可通过 `MaxConnections`、`MaxConcurrentRequests` 限制内存峰值。它仍通过标准 HTTP 请求/响应接口连接业务层，适合在现有 Handler / Service 上做网络层验证。
 
 业务代码可以直接使用 `git.yuntuo.hk/k2safe/go-omni/logx` 打结构化日志，不需要依赖具体网络库：
 

ROADMAP.md

@@ -42,6 +42,7 @@
 - 标准库与 native 共用 TCP benchmark，固定 1 / 32 / 128 并发；相同应用、客户端、超时和响应检查，避免两个测试入口逐渐失去可比性。
 - 新增独立服务进程压测 `internal/perfcheck`，记录成功吞吐、P50/P95/P99、服务端分配、采样堆峰值，以及 128 / 1,024 条空闲长连接的堆内存、goroutine 和断开后的回收情况。原始报告保存在 `performance/tcpBaseline.json`。
 - native 已支持受 `MaxBodyBytes` 约束的 `Content-Length` / chunked 请求体，仍使用标准 `http.Request` / `http.ResponseWriter` 作为桥接；解析器、连接抽象和生产压测将在后续阶段继续下沉。
+- native 响应已支持标准 `http.Flusher`；HTTP/1.1 流式响应使用 chunked framing，普通响应继续走 `Content-Length`，HTTP/1.0 回退到关闭连接分帧。
 - 已补充慢请求体超时、大请求体、HTTP/1.0 Keep-Alive、HEAD/204 和管线请求验收；Linux、慢客户端规模、TLS 和持续运行压测仍待在目标环境验证。
 - 已验证 `MaxConnections` 满载时快速拒绝新连接、请求完成后释放容量；后续仍需在目标部署环境观察文件描述符、内核 socket 缓冲和反向代理联动。
 - 已用 8 个并发 worker 完成 800 次建连/断连 race 验收；真实部署仍需补文件描述符和跨平台 socket 行为观测。

transport/README.md

@@ -16,6 +16,8 @@ registerRoutes(application) // 与 omni.New() 复用同一份业务注册代码
 
 `transport/native` 是第一版自研 HTTP/1.1 服务端：它自己管理 TCP listener、连接、Keep-Alive、请求解析、响应写回和优雅关闭，再把请求交给现有 `http.Handler`。请求体支持 `Content-Length` 和 chunked，并统一受 `MaxBodyBytes` 限制；其他传输编码会被拒绝。它仍使用标准 `http.Request` / `http.ResponseWriter` 作为业务桥接，因此不会影响现有 Handler / Service。
 
+业务 Handler 需要流式返回时可以断言 `http.Flusher` 并调用 `Flush()`；HTTP/1.1 会自动使用 chunked 响应，普通未 Flush 的响应仍自动生成 `Content-Length`。HTTP/1.0 没有 chunked framing，会退回到写完后关闭连接。
+
 服务端不是把连接交给一个全局串行 worker，而是每条连接独立处理请求；Keep-Alive 负责复用客户端连接，`Server` 内部的 `sync.Pool` 负责复用小型 Reader、响应对象和连接状态。Reader 默认只有 4 KiB，需要处理大请求头时按需增长。需要保护内存上限时，可以设置 `MaxConnections` 和 `MaxConcurrentRequests`；超出连接上限的连接会被关闭，超出请求并发上限的请求返回 503，不会无限排队。
 
 这个设计借鉴了 Nginx 的资源边界，但没有照搬它的进程模型：Nginx 用少量 worker 进程和事件机制管理大量连接，并用 `worker_connections` 限制每个 worker 的同时连接数；Go runtime 自己已经用 epoll / kqueue 等网络轮询器唤醒阻塞的 goroutine，因此 native 采用“accept loop + 每连接轻量 goroutine + 请求并发闸门”的等价实现。这样业务 Handler 可以继续执行同步代码，慢请求不会阻塞其他连接；只有在真实压测证明 goroutine 调度或协议状态机成为瓶颈时，才需要继续下沉为手写事件循环。

transport/native/server.go

@@ -334,6 +334,7 @@ func (s *Server) serveConnection(baseCtx context.Context, state *connection) {
 		state.active.Store(true)
 		writer := s.writerPool.Get().(*responseWriter)
 		writer.reset(request.Method)
+		writer.configure(request, state.conn, closeAfter, s.WriteTimeout)
 		requestSlot := false
 		if slots := s.requestSlots; slots != nil {
 			select {
@@ -378,10 +379,11 @@ func (s *Server) serveConnection(baseCtx context.Context, state *connection) {
 		if requestSlot {
 			<-s.requestSlots
 		}
+		shouldClose := closeAfter || writer.closeAfter || closing
 		writer.reset("")
 		s.writerPool.Put(writer)
 		state.active.Store(false)
-		if closeAfter || closing {
+		if shouldClose {
 			return
 		}
 	}
@@ -719,11 +721,20 @@ func writeSimple(conn net.Conn, status int, body string, closeAfter bool, timeou
 }
 
 type responseWriter struct {
-	header        http.Header
-	status        int
-	wroteHeader   bool
-	body          bytes.Buffer
-	requestMethod string
+	header         http.Header
+	status         int
+	wroteHeader    bool
+	body           bytes.Buffer
+	requestMethod  string
+	protoMajor     int
+	protoMinor     int
+	conn           net.Conn
+	closeAfter     bool
+	writeTimeout   time.Duration
+	streaming      bool
+	chunked        bool
+	streamingClose bool
+	writeErr       error
 }
 
 func newResponseWriter(method string) *responseWriter {
@@ -742,6 +753,23 @@ func (w *responseWriter) reset(method string) {
 		w.body.Reset()
 	}
 	w.requestMethod = method
+	w.protoMajor = 0
+	w.protoMinor = 0
+	w.conn = nil
+	w.closeAfter = false
+	w.writeTimeout = 0
+	w.streaming = false
+	w.chunked = false
+	w.streamingClose = false
+	w.writeErr = nil
+}
+
+func (w *responseWriter) configure(request *http.Request, conn net.Conn, closeAfter bool, timeout time.Duration) {
+	w.protoMajor = request.ProtoMajor
+	w.protoMinor = request.ProtoMinor
+	w.conn = conn
+	w.closeAfter = closeAfter
+	w.writeTimeout = timeout
 }
 
 func (w *responseWriter) Header() http.Header { return w.header }
@@ -755,13 +783,121 @@ func (w *responseWriter) WriteHeader(status int) {
 }
 
 func (w *responseWriter) Write(data []byte) (int, error) {
+	if w.writeErr != nil {
+		return 0, w.writeErr
+	}
 	if !w.wroteHeader {
 		w.WriteHeader(http.StatusOK)
 	}
+	if w.streaming {
+		written, err := w.writeStream(data)
+		if err != nil {
+			w.writeErr = err
+		}
+		return written, err
+	}
 	return w.body.Write(data)
 }
 
+// Flush implements http.Flusher. Once a response is flushed, HTTP/1.1 uses
+// chunked framing unless the handler supplied Content-Length. HTTP/1.0 falls
+// back to close-delimited output because it has no chunked framing.
+func (w *responseWriter) Flush() {
+	if w.writeErr != nil || w.streaming || w.requestMethod == http.MethodHead || w.status == http.StatusNoContent ||
+		w.status == http.StatusResetContent || w.status == http.StatusNotModified || (w.status >= 100 && w.status < 200) {
+		return
+	}
+	if err := w.beginStream(); err != nil {
+		w.writeErr = err
+	}
+}
+
+func (w *responseWriter) beginStream() error {
+	if w.conn == nil {
+		return errors.New("go-omni/native: response stream is not connected")
+	}
+	if !w.wroteHeader {
+		w.status = http.StatusOK
+		w.wroteHeader = true
+	}
+	if w.requestMethod == http.MethodHead || w.status == http.StatusNoContent || w.status == http.StatusResetContent ||
+		w.status == http.StatusNotModified || (w.status >= 100 && w.status < 200) {
+		return nil
+	}
+	contentLength := w.header.Get("Content-Length")
+	w.chunked = contentLength == "" && w.protoMajor == 1 && w.protoMinor == 1
+	if w.chunked {
+		w.header.Set("Transfer-Encoding", "chunked")
+	} else if contentLength == "" {
+		w.streamingClose = true
+	}
+	closeAfter := w.closeAfter || w.streamingClose
+	if closeAfter {
+		w.header.Set("Connection", "close")
+	}
+	w.closeAfter = closeAfter
+	if err := w.writeHeaders(w.conn); err != nil {
+		return err
+	}
+	w.streaming = true
+	buffered := w.body.Bytes()
+	w.body.Reset()
+	if len(buffered) == 0 {
+		return nil
+	}
+	_, err := w.writeStream(buffered)
+	return err
+}
+
+func (w *responseWriter) writeStream(data []byte) (int, error) {
+	if len(data) == 0 {
+		return 0, nil
+	}
+	if w.conn == nil {
+		return 0, errors.New("go-omni/native: response stream is not connected")
+	}
+	if w.chunked {
+		var prefix [32]byte
+		framed := strconv.AppendInt(prefix[:0], int64(len(data)), 16)
+		framed = append(framed, '\r', '\n')
+		if err := w.writeBytes(framed); err != nil {
+			return 0, err
+		}
+		if err := w.writeBytes(data); err != nil {
+			return 0, err
+		}
+		if err := w.writeBytes([]byte("\r\n")); err != nil {
+			return 0, err
+		}
+		return len(data), nil
+	}
+	if err := w.writeBytes(data); err != nil {
+		return 0, err
+	}
+	return len(data), nil
+}
+
+func (w *responseWriter) writeBytes(data []byte) error {
+	if w.writeTimeout > 0 {
+		_ = w.conn.SetWriteDeadline(time.Now().Add(w.writeTimeout))
+	}
+	return writeAll(w.conn, data)
+}
+
 func (w *responseWriter) writeTo(conn net.Conn, closeAfter bool, timeout time.Duration) error {
+	if w.streaming {
+		if w.writeErr != nil {
+			_ = conn.SetWriteDeadline(time.Time{})
+			return w.writeErr
+		}
+		if w.chunked {
+			if err := w.writeBytes([]byte("0\r\n\r\n")); err != nil {
+				w.writeErr = err
+			}
+		}
+		_ = conn.SetWriteDeadline(time.Time{})
+		return w.writeErr
+	}
 	if timeout > 0 {
 		_ = conn.SetWriteDeadline(time.Now().Add(timeout))
 	}
@@ -775,7 +911,7 @@ func (w *responseWriter) writeTo(conn net.Conn, closeAfter bool, timeout time.Du
 	if noBody {
 		body = nil
 	}
-	if w.header.Get("Content-Length") == "" && w.status >= 200 && w.status != http.StatusNoContent && w.status != http.StatusNotModified && w.status != http.StatusResetContent {
+	if w.header.Get("Content-Length") == "" && w.header.Get("Transfer-Encoding") == "" && w.status >= 200 && w.status != http.StatusNoContent && w.status != http.StatusNotModified && w.status != http.StatusResetContent {
 		w.header.Set("Content-Length", strconv.Itoa(len(w.body.Bytes())))
 	}
 	if closeAfter {
@@ -808,7 +944,56 @@ func (w *responseWriter) writeTo(conn net.Conn, closeAfter bool, timeout time.Du
 	}
 	output.WriteString("\r\n")
 	output.Write(body)
-	_, err := conn.Write(output.Bytes())
+	err := writeAll(conn, output.Bytes())
 	_ = conn.SetWriteDeadline(time.Time{})
 	return err
 }
+
+func (w *responseWriter) writeHeaders(conn net.Conn) error {
+	if w.writeTimeout > 0 {
+		_ = conn.SetWriteDeadline(time.Now().Add(w.writeTimeout))
+	}
+	keys := make([]string, 0, len(w.header))
+	for key := range w.header {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	var output bytes.Buffer
+	output.WriteString("HTTP/1.1 ")
+	output.WriteString(strconv.Itoa(w.status))
+	output.WriteByte(' ')
+	output.WriteString(http.StatusText(w.status))
+	output.WriteString("\r\n")
+	for _, key := range keys {
+		if !validToken(key) {
+			return errors.New("invalid response header name")
+		}
+		for _, value := range w.header.Values(key) {
+			if !validHeaderValue(value) {
+				return errors.New("invalid response header")
+			}
+			output.WriteString(key)
+			output.WriteString(": ")
+			output.WriteString(value)
+			output.WriteString("\r\n")
+		}
+	}
+	output.WriteString("\r\n")
+	return writeAll(conn, output.Bytes())
+}
+
+func writeAll(conn net.Conn, data []byte) error {
+	for len(data) > 0 {
+		written, err := conn.Write(data)
+		if written > 0 {
+			data = data[written:]
+		}
+		if err != nil {
+			return err
+		}
+		if written == 0 {
+			return io.ErrShortWrite
+		}
+	}
+	return nil
+}

transport/native/server_test.go

@@ -225,6 +225,60 @@ func TestServerPipelinedRequestsAndLargeBody(t *testing.T) {
 	}
 }
 
+func TestServerFlushesChunkedResponses(t *testing.T) {
+	flushed := make(chan struct{})
+	release := make(chan struct{})
+	application := omni.NewWithRouter(netrouter.New())
+	application.GET("/stream", func(ctx *omni.Context) error {
+		flusher, ok := ctx.Raw().(http.Flusher)
+		if !ok {
+			return fmt.Errorf("native response writer does not implement http.Flusher")
+		}
+		ctx.Raw().Header().Set("Content-Type", "text/plain")
+		_, _ = io.WriteString(ctx.Raw(), "one\n")
+		flusher.Flush()
+		close(flushed)
+		<-release
+		_, _ = io.WriteString(ctx.Raw(), "two\n")
+		return nil
+	})
+	server, listener, done := startServer(t, application)
+	defer stopServer(t, server, listener, done)
+
+	conn, err := net.Dial("tcp", listener.Addr().String())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer conn.Close()
+	reader := bufio.NewReader(conn)
+	_, _ = io.WriteString(conn, "GET /stream HTTP/1.1\r\nHost: test\r\nConnection: close\r\n\r\n")
+	response, err := http.ReadResponse(reader, &http.Request{Method: http.MethodGet})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if response.StatusCode != http.StatusOK || response.TransferEncoding == nil || response.TransferEncoding[0] != "chunked" {
+		response.Body.Close()
+		t.Fatalf("expected chunked response, status=%d transfer=%v", response.StatusCode, response.TransferEncoding)
+	}
+	select {
+	case <-flushed:
+	case <-time.After(time.Second):
+		response.Body.Close()
+		t.Fatal("handler did not flush response")
+	}
+	first := make([]byte, len("one\n"))
+	if _, err := io.ReadFull(response.Body, first); err != nil || string(first) != "one\n" {
+		response.Body.Close()
+		t.Fatalf("unexpected first stream chunk: %q error=%v", first, err)
+	}
+	close(release)
+	rest, err := io.ReadAll(response.Body)
+	response.Body.Close()
+	if err != nil || string(rest) != "two\n" {
+		t.Fatalf("unexpected remaining stream body: %q error=%v", rest, err)
+	}
+}
+
 func TestServerProtocolBoundaries(t *testing.T) {
 	application := omni.NewWithRouter(netrouter.New())
 	application.GET("/health", func(ctx *omni.Context) error { return ctx.OK("ok") })