Nagle's algorithm adds latency in order to reduce network packet overhead on small packets. But when we are already using corking to merge smaller packets into transactional requests, the extra delay from TCP defaults just gets in the way (see recent commit bd2cd4a4).
For reference, qemu as an NBD client already requests TCP_NODELAY (see nbd_connect() in nbd/client-connection.c); as does libnbd as a client [1], and nbdkit as a server [2]. Furthermore, the NBD spec recommends the use of TCP_NODELAY [3]. [1] https://gitlab.com/nbdkit/libnbd/-/blob/a48a1142/generator/states-connect.c#L39 [2] https://gitlab.com/nbdkit/nbdkit/-/blob/45b72f5b/server/sockets.c#L430 [3] https://github.com/NetworkBlockDevice/nbd/blob/master/doc/proto.md#protocol-phases CC: Florian Westphal <f...@strlen.de> Signed-off-by: Eric Blake <ebl...@redhat.com> Message-Id: <20230404004047.142086-1-ebl...@redhat.com> Reviewed-by: Philippe Mathieu-Daudé <phi...@linaro.org> --- nbd/server.c | 1 + 1 file changed, 1 insertion(+) diff --git a/nbd/server.c b/nbd/server.c index 848836d4140..3d8d0d81df2 100644 --- a/nbd/server.c +++ b/nbd/server.c @@ -2758,6 +2758,7 @@ void nbd_client_new(QIOChannelSocket *sioc, } client->tlsauthz = g_strdup(tlsauthz); client->sioc = sioc; + qio_channel_set_delay(QIO_CHANNEL(sioc), false); object_ref(OBJECT(client->sioc)); client->ioc = QIO_CHANNEL(sioc); object_ref(OBJECT(client->ioc)); -- 2.39.2
