From: Vladimír Čunát <vladimir.cunat@nic.cz>
Date: Mon, 10 Jun 2024 13:03:23 +0000 (+0200)
Subject: daemon: set options on sockets towards clients (optionally)
X-Git-Tag: v5.7.4^2~3^2~2
X-Git-Url: http://git.ipfire.org/gitweb/gitweb.cgi?a=commitdiff_plain;h=refs%2Fenvironments%2Fdocs-develop-tmp-xva6ir%2Fdeployments%2F4625;p=thirdparty%2Fknot-resolver.git

daemon: set options on sockets towards clients (optionally)
---

diff --git a/daemon/io.c b/daemon/io.c
index 48ee96bba..5d4e7f9a5 100644
--- a/daemon/io.c
+++ b/daemon/io.c
@@ -251,6 +251,17 @@ int io_bind(const struct sockaddr *addr, int type, const endpoint_flags_t *flags
 	return fd;
 }
 
+/// Optionally set a socket option and log error on failure.
+static void set_so(int fd, int so_option, int value, const char *descr)
+{
+	if (!value) return;
+	if (setsockopt(fd, SOL_SOCKET, so_option, &value, sizeof(value))) {
+		kr_log_error(IO, "failed to set %s to %d: %s\n",
+				descr, value, strerror(errno));
+		// we treat this as non-critical failure
+	}
+}
+
 int io_listen_udp(uv_loop_t *loop, uv_udp_t *handle, int fd)
 {
 	if (!handle) {
@@ -262,6 +273,10 @@ int io_listen_udp(uv_loop_t *loop, uv_udp_t *handle, int fd)
 	ret = uv_udp_open(handle, fd);
 	if (ret) return ret;
 
+	struct network *net = &the_worker->engine->net;
+	set_so(fd, SO_SNDBUF, net->listen_udp_buflens.snd, "UDP send buffer size");
+	set_so(fd, SO_RCVBUF, net->listen_udp_buflens.rcv, "UDP receive buffer size");
+
 	uv_handle_t *h = (uv_handle_t *)handle;
 	check_bufsize(h);
 	/* Handle is already created, just create context. */
@@ -702,6 +717,18 @@ int io_listen_tcp(uv_loop_t *loop, uv_tcp_t *handle, int fd, int tcp_backlog, bo
 	}
 #endif
 
+	/* These get inherited into the individual connections (on Linux at least). */
+	struct network *net = &the_worker->engine->net;
+	set_so(fd, SO_SNDBUF, net->listen_tcp_buflens.snd, "TCP send buffer size");
+	set_so(fd, SO_RCVBUF, net->listen_tcp_buflens.rcv, "TCP receive buffer size");
+#ifdef TCP_USER_TIMEOUT
+	val = net->tcp.user_timeout;
+	if (val && setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &val, sizeof(val))) {
+		kr_log_error(IO, "listen TCP (user_timeout): %s\n", strerror(errno));
+	}
+	// TODO: also for upstream connections, at least this one option?
+#endif
+
 	handle->data = NULL;
 	return 0;
 }
diff --git a/daemon/network.c b/daemon/network.c
index a20b1e453..598b1dcd4 100644
--- a/daemon/network.c
+++ b/daemon/network.c
@@ -73,6 +73,11 @@ void network_init(struct network *net, uv_loop_t *loop, int tcp_backlog)
 		net->tcp.in_idle_timeout = 10000;
 		net->tcp.tls_handshake_timeout = TLS_MAX_HANDSHAKE_TIME;
 		net->tcp_backlog = tcp_backlog;
+		net->tcp.user_timeout = 1000; // 1s should be more than enough
+
+		// On Linux, unset means some auto-tuning mechanism also depending on RAM,
+		// which might be OK default (together with the user_timeout above)
+		//net->listen_{tcp,udp}_buflens.{snd,rcv}
 	}
 }
 
diff --git a/daemon/network.h b/daemon/network.h
index e21651fdd..c9cd621bb 100644
--- a/daemon/network.h
+++ b/daemon/network.h
@@ -67,6 +67,10 @@ typedef array_t(struct endpoint) endpoint_array_t;
 struct net_tcp_param {
 	uint64_t in_idle_timeout;
 	uint64_t tls_handshake_timeout;
+
+	/** Milliseconds of unacknowledged data; see TCP_USER_TIMEOUT in man tcp.7
+	 * Linux only, probably. */
+	unsigned int user_timeout;
 };
 
 /** Information about an address that is allowed to use PROXYv2. */
@@ -104,6 +108,13 @@ struct network {
 	struct tls_session_ticket_ctx *tls_session_ticket_ctx;
 	struct net_tcp_param tcp;
 	int tcp_backlog;
+
+	/** Kernel-side buffer sizes for sending and receiving. (in bytes)
+	 * They are per socket, so in the TCP case they are per connection.
+	 * See SO_SNDBUF and SO_RCVBUF in man socket.7  These are in POSIX. */
+	struct {
+		int snd, rcv;
+	} listen_udp_buflens, listen_tcp_buflens;
 };
 
 void network_init(struct network *net, uv_loop_t *loop, int tcp_backlog);