]> git.proxmox.com Git - mirror_qemu.git/commitdiff
nbd/server: Request TCP_NODELAY
authorEric Blake <eblake@redhat.com>
Tue, 4 Apr 2023 00:40:47 +0000 (19:40 -0500)
committerEric Blake <eblake@redhat.com>
Tue, 4 Apr 2023 13:13:15 +0000 (08:13 -0500)
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 <fw@strlen.de>
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <20230404004047.142086-1-eblake@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
nbd/server.c

index 848836d41405ac3ddf9ae20af5cafd4de0b676d8..3d8d0d81df2295cd196f15d0189184422402862a 100644 (file)
@@ -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));