From: Remi Gacogne <remi.gacogne@powerdns.com>
Date: Thu, 30 Jun 2022 09:06:24 +0000 (+0200)
Subject: dnsdist: Set TCP_NODELAY on the TCP connection to backends
X-Git-Tag: auth-4.8.0-alpha0~23^2
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=059c8f524ee1e467c06787d591b779e08a279439;p=thirdparty%2Fpdns.git

dnsdist: Set TCP_NODELAY on the TCP connection to backends

Setting `TCP_NODELAY` disables Nagle's algorithm, which is unfortunately
not always playing nice with TCP delayed ACKs. This is especially true
when the backend supports out-of-order processing, allowing dnsdist to
send several queries at once without waiting for an answer. In that case
dnsdist often has to wait several milliseconds (20ms is common) before
the remote end decides to send the delayed ACK which allows us to send
the next query.
---

diff --git a/pdns/dnsdistdist/dnsdist-tcp-downstream.cc b/pdns/dnsdistdist/dnsdist-tcp-downstream.cc
index 418bf6b5f0..72809fbbe7 100644
--- a/pdns/dnsdistdist/dnsdist-tcp-downstream.cc
+++ b/pdns/dnsdistdist/dnsdist-tcp-downstream.cc
@@ -76,6 +76,12 @@ bool ConnectionToBackend::reconnect()
       auto socket = std::make_unique<Socket>(d_ds->d_config.remote.sin4.sin_family, SOCK_STREAM, 0);
       DEBUGLOG("result of socket() is "<<socket->getHandle());
 
+      /* disable NAGLE, which does not play nicely with delayed ACKs.
+         In theory we could be wasting up to 500 milliseconds waiting for
+         the other end to acknowledge our initial packet before we could
+         send the rest. */
+      setTCPNoDelay(socket->getHandle());
+
 #ifdef SO_BINDTODEVICE
       if (!d_ds->d_config.sourceItfName.empty()) {
         int res = setsockopt(socket->getHandle(), SOL_SOCKET, SO_BINDTODEVICE, d_ds->d_config.sourceItfName.c_str(), d_ds->d_config.sourceItfName.length());