grpc: disable and document overrides of OS default TCP keepalive by Go (#6672)

Co-authored-by: Arvind Bright <arvind.bright100@gmail.com>

benchmark/benchmain/main.go

@@ -374,7 +374,7 @@ func makeClients(bf stats.Features) ([]testgrpc.BenchmarkServiceClient, func())
 			logger.Fatalf("Failed to listen: %v", err)
 		}
 		opts = append(opts, grpc.WithContextDialer(func(ctx context.Context, address string) (net.Conn, error) {
-			return nw.ContextDialer((&net.Dialer{}).DialContext)(ctx, "tcp", lis.Addr().String())
+			return nw.ContextDialer((&net.Dialer{KeepAlive: time.Duration(-1)}).DialContext)(ctx, "tcp", lis.Addr().String())
 		}))
 	}
 	lis = nw.Listener(lis)

dialoptions.go

@@ -414,6 +414,15 @@ func WithTimeout(d time.Duration) DialOption {
 // connections. If FailOnNonTempDialError() is set to true, and an error is
 // returned by f, gRPC checks the error's Temporary() method to decide if it
 // should try to reconnect to the network address.
+//
+// Note: As of Go 1.21, the standard library overrides the OS defaults for
+// TCP keepalive time and interval to 15s.
+// To retain OS defaults, use a net.Dialer with the KeepAlive field set to a
+// negative value.
+//
+// For more information, please see [issue 23459] in the Go github repo.
+//
+// [issue 23459]: https://github.com/golang/go/issues/23459
 func WithContextDialer(f func(context.Context, string) (net.Conn, error)) DialOption {
 	return newFuncDialOption(func(o *dialOptions) {
 		o.copts.Dialer = f

internal/transport/http2_client.go

@@ -176,7 +176,9 @@ func dial(ctx context.Context, fn func(context.Context, string) (net.Conn, error
 	if networkType == "tcp" && useProxy {
 		return proxyDial(ctx, address, grpcUA)
 	}
-	return (&net.Dialer{}).DialContext(ctx, networkType, address)
+	// KeepAlive is set to a negative value to prevent Go's override of the TCP
+	// keepalive time and interval; retain the OS default.
+	return (&net.Dialer{KeepAlive: time.Duration(-1)}).DialContext(ctx, networkType, address)
 }
 
 func isTemporary(err error) bool {

internal/transport/proxy.go

@@ -28,6 +28,7 @@ import (
 	"net/http"
 	"net/http/httputil"
 	"net/url"
+	"time"
 )
 
 const proxyAuthHeaderKey = "Proxy-Authorization"
@@ -122,7 +123,7 @@ func proxyDial(ctx context.Context, addr string, grpcUA string) (conn net.Conn,
 		newAddr = proxyURL.Host
 	}
 
-	conn, err = (&net.Dialer{}).DialContext(ctx, "tcp", newAddr)
+	conn, err = (&net.Dialer{KeepAlive: time.Duration(-1)}).DialContext(ctx, "tcp", newAddr)
 	if err != nil {
 		return
 	}

server.go

@@ -813,6 +813,15 @@ func (l *listenSocket) Close() error {
 // Serve returns when lis.Accept fails with fatal errors.  lis will be closed when
 // this method returns.
 // Serve will return a non-nil error unless Stop or GracefulStop is called.
+//
+// Note: As of Go 1.21, the standard library overrides the OS defaults for
+// TCP keepalive time and interval to 15s.
+// To retain OS defaults, pass a net.Listener created by calling the Listen method
+// on a net.ListenConfig with the `KeepAlive` field set to a negative value.
+//
+// For more information, please see [issue 23459] in the Go github repo.
+//
+// [issue 23459]: https://github.com/golang/go/issues/23459
 func (s *Server) Serve(lis net.Listener) error {
 	s.mu.Lock()
 	s.printf("serving")