qemu-kvm/SOURCES/kvm-nbd-server-Request-TCP_NODELAY.patch

56 lines
2.2 KiB
Diff

From 17c5524ada3f2ca9a9c645f540bedc5575302059 Mon Sep 17 00:00:00 2001
From: Eric Blake <eblake@redhat.com>
Date: Mon, 3 Apr 2023 19:40:47 -0500
Subject: [PATCH 5/5] nbd/server: Request TCP_NODELAY
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
RH-Author: Eric Blake <eblake@redhat.com>
RH-MergeRequest: 274: nbd: improve TLS performance of NBD server
RH-Bugzilla: 2035712
RH-Acked-by: Miroslav Rezanina <mrezanin@redhat.com>
RH-Acked-by: Kevin Wolf <kwolf@redhat.com>
RH-Acked-by: Stefano Garzarella <sgarzare@redhat.com>
RH-Commit: [2/2] 092145077756cda2a4f849c5911031b0fc4a2134 (ebblake/qemu-kvm)
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>
(cherry picked from commit f1426881a827a6d3f31b65616c4a8db1e9e7c45e)
Signed-off-by: Eric Blake <eblake@redhat.com>
---
nbd/server.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/nbd/server.c b/nbd/server.c
index a5edc7f681..6db124cf53 100644
--- a/nbd/server.c
+++ b/nbd/server.c
@@ -2738,6 +2738,7 @@ 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));
--
2.39.1