net: tcp: Implement Nagle's algorithm

To improve the performance with small chunks send, implement Nagle's
algorithm. Provide the option TCP_NODELAY to disable the algorithm.

Signed-off-by: Sjors Hettinga <s.a.hettinga@gmail.com>

subsys/net/ip/tcp.c

@@ -848,6 +848,37 @@ static int ip_header_add(struct tcp *conn, struct net_pkt *pkt)
 	return -EINVAL;
 }
 
+static int set_tcp_nodelay(struct tcp *conn, const void *value, size_t len)
+{
+	int no_delay_int;
+
+	if (len != sizeof(int)) {
+		return -EINVAL;
+	}
+
+	no_delay_int = *(int *)value;
+
+	if ((no_delay_int < 0) || (no_delay_int > 1)) {
+		return -EINVAL;
+	}
+
+	conn->tcp_nodelay = (bool)no_delay_int;
+
+	return 0;
+}
+
+static int get_tcp_nodelay(struct tcp *conn, void *value, size_t *len)
+{
+	int no_delay_int = (int)conn->tcp_nodelay;
+
+	*((int *)value) = no_delay_int;
+
+	if (len) {
+		*len = sizeof(int);
+	}
+	return 0;
+}
+
 static int net_tcp_set_mss_opt(struct tcp *conn, struct net_pkt *pkt)
 {
 	NET_PKT_DATA_ACCESS_DEFINE(mss_opt_access, struct tcp_mss_option);
@@ -1097,6 +1128,19 @@ static int tcp_send_queued_data(struct tcp *conn)
 	}
 
 	while (tcp_unsent_len(conn) > 0) {
+		/* Implement Nagle's algorithm */
+		if ((conn->tcp_nodelay == false) && (conn->unacked_len > 0)) {
+			/* If there is already pending data */
+			if (tcp_unsent_len(conn) < conn_mss(conn)) {
+				/* The number of bytes to be transmitted is less than an MSS,
+				 * skip transmission for now.
+				 * Wait for more data to be transmitted or all pending data
+				 * being acknowledged.
+				 */
+				break;
+			}
+		}
+
 		ret = tcp_send_data(conn);
 		if (ret < 0) {
 			break;
@@ -1301,6 +1345,7 @@ static struct tcp *tcp_conn_alloc(struct net_context *context)
 	conn->in_connect = false;
 	conn->state = TCP_LISTEN;
 	conn->recv_win = tcp_window;
+	conn->tcp_nodelay = false;
 
 	/* Set the recv_win with the rcvbuf configured for the socket. */
 	if (IS_ENABLED(CONFIG_NET_CONTEXT_RCVBUF) &&
@@ -2994,6 +3039,56 @@ uint16_t net_tcp_get_supported_mss(const struct tcp *conn)
 	return 0;
 }
 
+int net_tcp_set_option(struct net_context *context,
+		       enum tcp_conn_option option,
+		       const void *value, size_t len)
+{
+	int ret = 0;
+
+	NET_ASSERT(context);
+
+	struct tcp *conn = context->tcp;
+
+	NET_ASSERT(conn);
+
+	k_mutex_lock(&conn->lock, K_FOREVER);
+
+	switch (option) {
+	case TCP_OPT_NODELAY:
+		ret = set_tcp_nodelay(conn, value, len);
+		break;
+	}
+
+	k_mutex_unlock(&conn->lock);
+
+	return ret;
+}
+
+int net_tcp_get_option(struct net_context *context,
+		       enum tcp_conn_option option,
+		       void *value, size_t *len)
+{
+	int ret = 0;
+
+	NET_ASSERT(context);
+
+	struct tcp *conn = context->tcp;
+
+	NET_ASSERT(conn);
+
+	k_mutex_lock(&conn->lock, K_FOREVER);
+
+	switch (option) {
+	case TCP_OPT_NODELAY:
+		ret = get_tcp_nodelay(conn, value, len);
+		break;
+	}
+
+	k_mutex_unlock(&conn->lock);
+
+	return ret;
+}
+
 const char *net_tcp_state_str(enum tcp_state state)
 {
 	return tcp_state_to_str(state, false);

subsys/net/ip/tcp_internal.h

@@ -30,6 +30,10 @@ extern "C" {
 
 #include "tcp_private.h"
 
+enum tcp_conn_option {
+	TCP_OPT_NODELAY	= 1,
+};
+
 /**
  * @brief Calculates and returns the MSS for a given TCP context
  *
@@ -367,6 +371,57 @@ void net_tcp_init(void);
 #define net_tcp_init(...)
 #endif
 
+/**
+ * @brief Set tcp specific options of a socket
+ *
+ * @param context Network context
+ *
+ * @return 0 on success, -EINVAL if the value is not allowed
+ */
+#if defined(CONFIG_NET_NATIVE_TCP)
+int net_tcp_set_option(struct net_context *context,
+		       enum tcp_conn_option option,
+		       const void *value, size_t len);
+#else
+static inline int net_tcp_set_option(struct net_context *context,
+				     enum tcp_conn_option option,
+				     const void *value, size_t len)
+{
+	ARG_UNUSED(context);
+	ARG_UNUSED(option);
+	ARG_UNUSED(value);
+	ARG_UNUSED(len);
+
+	return -EPROTONOSUPPORT;
+}
+#endif
+
+
+/**
+ * @brief Obtain tcp specific options of a socket
+ *
+ * @param context Network context
+ *
+ * @return 0 on success
+ */
+#if defined(CONFIG_NET_NATIVE_TCP)
+int net_tcp_get_option(struct net_context *context,
+		       enum tcp_conn_option option,
+		       void *value, size_t *len);
+#else
+static inline int net_tcp_get_option(struct net_context *context,
+				     enum tcp_conn_option option,
+				     void *value, size_t *len)
+{
+	ARG_UNUSED(context);
+	ARG_UNUSED(option);
+	ARG_UNUSED(value);
+	ARG_UNUSED(len);
+
+	return -EPROTONOSUPPORT;
+}
+#endif
+
 /**
  * @brief Obtain a semaphore indicating if transfers are blocked (either due to
  *        filling TX window or entering retransmission mode).

subsys/net/ip/tcp_private.h

@@ -269,6 +269,7 @@ struct tcp { /* TCP connection */
 	bool in_retransmission : 1;
 	bool in_connect : 1;
 	bool in_close : 1;
+	bool tcp_nodelay : 1;
 };
 
 #define _flags(_fl, _op, _mask, _cond)					\

subsys/net/lib/sockets/sockets.c

@@ -1852,6 +1852,12 @@ int zsock_getsockopt_ctx(struct net_context *ctx, int level, int optname,
 			}
 			break;
 		}
+	case IPPROTO_TCP:
+		switch (optname) {
+		case TCP_NODELAY:
+			ret = net_tcp_get_option(ctx, TCP_OPT_NODELAY, optval, optlen);
+			return ret;
+		}
 	}
 
 	errno = ENOPROTOOPT;
@@ -2099,10 +2105,9 @@ int zsock_setsockopt_ctx(struct net_context *ctx, int level, int optname,
 	case IPPROTO_TCP:
 		switch (optname) {
 		case TCP_NODELAY:
-			/* Ignore for now. Provided to let port
-			 * existing apps.
-			 */
-			return 0;
+			ret = net_tcp_set_option(ctx,
+						 TCP_OPT_NODELAY, optval, optlen);
+			return ret;
 		}
 		break;