XDP-TCP: announce correct MSS

src/libknot/xdp/bpf-user.h

@@ -77,6 +77,9 @@ struct knot_xdp_socket {
 
 	/*! The kernel has to be woken up by a syscall indication. */
 	bool kernel_needs_wakeup;
+
+	/*! The limit of frame size. */
+	unsigned frame_limit;
 };
 
 /*!

src/libknot/xdp/msg_init.h

@@ -45,12 +45,15 @@ inline static void msg_init(knot_xdp_msg_t *msg, knot_xdp_msg_flag_t flags)
 	if (flags & KNOT_XDP_MSG_TCP) {
 		msg->ackno = 0;
 		msg->seqno = dnssec_random_uint32_t();
+		if (flags & KNOT_XDP_MSG_SYN) {
+			msg->flags |= KNOT_XDP_MSG_MSS;
+		}
 	}
 }
 
 inline static void msg_init_reply(knot_xdp_msg_t *msg, const knot_xdp_msg_t *query)
 {
-	msg_init_base(msg, query->flags & (KNOT_XDP_MSG_IPV6 | KNOT_XDP_MSG_TCP));
+	msg_init_base(msg, query->flags & (KNOT_XDP_MSG_IPV6 | KNOT_XDP_MSG_TCP | KNOT_XDP_MSG_MSS));
 
 	memcpy(msg->eth_from, query->eth_to,   ETH_ALEN);
 	memcpy(msg->eth_to,   query->eth_from, ETH_ALEN);

src/libknot/xdp/protocols.h

@@ -262,7 +262,8 @@ inline static void prot_write_udp(void *data, const knot_xdp_msg_t *msg, void *d
 }
 
 inline static void prot_write_tcp(void *data, const knot_xdp_msg_t *msg, void *data_end,
-                                  uint16_t src_port, uint16_t dst_port, uint32_t chksum)
+                                  uint16_t src_port, uint16_t dst_port, uint32_t chksum,
+                                  uint16_t mss)
 {
 	struct tcphdr *tcp = data;
 
@@ -285,7 +286,7 @@ inline static void prot_write_tcp(void *data, const knot_xdp_msg_t *msg, void *d
 	hdr_end += PROT_TCP_OPT_LEN_WSC;
 	*hdr_end++ = PROT_TCP_OPT_NOOP;
 	if (msg->flags & KNOT_XDP_MSG_MSS) {
-		uint16_t mss = htobe16(1460); // TODO: set proper MSS value
+		mss = htobe16(mss);
 		hdr_end[0] = PROT_TCP_OPT_MSS;
 		hdr_end[1] = PROT_TCP_OPT_LEN_MSS;
 		memcpy(&hdr_end[2], &mss, sizeof(mss));
@@ -322,7 +323,8 @@ inline static uint16_t ipv4_checksum(const uint16_t *ipv4_hdr)
 	return ~from32to16(sum32);
 }
 
-inline static void prot_write_ipv4(void *data, const knot_xdp_msg_t *msg, void *data_end)
+inline static void prot_write_ipv4(void *data, const knot_xdp_msg_t *msg,
+                                   void *data_end, uint16_t tcp_mss)
 {
 	struct iphdr *ip4 = data;
 
@@ -349,13 +351,14 @@ inline static void prot_write_ipv4(void *data, const knot_xdp_msg_t *msg, void *
 		checksum(&chk, &src->sin_addr, sizeof(src->sin_addr));
 		checksum(&chk, &dst->sin_addr, sizeof(dst->sin_addr));
 
-		prot_write_tcp(data, msg, data_end, src->sin_port, dst->sin_port, chk);
+		prot_write_tcp(data, msg, data_end, src->sin_port, dst->sin_port, chk, tcp_mss);
 	} else {
 		prot_write_udp(data, msg, data_end, src->sin_port, dst->sin_port, 0); // IPv4/UDP requires no checksum
 	}
 }
 
-inline static void prot_write_ipv6(void *data, const knot_xdp_msg_t *msg, void *data_end)
+inline static void prot_write_ipv6(void *data, const knot_xdp_msg_t *msg,
+                                   void *data_end, uint16_t tcp_mss)
 {
 	struct ipv6hdr *ip6 = data;
 
@@ -379,13 +382,14 @@ inline static void prot_write_ipv6(void *data, const knot_xdp_msg_t *msg, void *
 	checksum(&chk, &dst->sin6_addr, sizeof(dst->sin6_addr));
 
 	if (msg->flags & KNOT_XDP_MSG_TCP) {
-		prot_write_tcp(data, msg, data_end, src->sin6_port, dst->sin6_port, chk);
+		prot_write_tcp(data, msg, data_end, src->sin6_port, dst->sin6_port, chk, tcp_mss);
 	} else {
 		prot_write_udp(data, msg, data_end, src->sin6_port, dst->sin6_port, chk);
 	}
 }
 
-inline static void prot_write_eth(void *data, const knot_xdp_msg_t *msg, void *data_end)
+inline static void prot_write_eth(void *data, const knot_xdp_msg_t *msg,
+                                  void *data_end, uint16_t tcp_mss)
 {
 	struct ethhdr *eth = data;
 
@@ -396,9 +400,9 @@ inline static void prot_write_eth(void *data, const knot_xdp_msg_t *msg, void *d
 
 	if (msg->flags & KNOT_XDP_MSG_IPV6) {
 		eth->h_proto = __constant_htons(ETH_P_IPV6);
-		prot_write_ipv6(data, msg, data_end);
+		prot_write_ipv6(data, msg, data_end, tcp_mss);
 	} else {
 		eth->h_proto = __constant_htons(ETH_P_IP);
-		prot_write_ipv4(data, msg, data_end);
+		prot_write_ipv4(data, msg, data_end, tcp_mss);
 	}
 }

src/libknot/xdp/xdp.c

@@ -29,6 +29,7 @@
 #include "libknot/endian.h"
 #include "libknot/errcode.h"
 #include "libknot/xdp/bpf-user.h"
+#include "libknot/xdp/eth.h"
 #include "libknot/xdp/msg_init.h"
 #include "libknot/xdp/protocols.h"
 #include "libknot/xdp/xdp.h"
@@ -179,6 +180,12 @@ int knot_xdp_init(knot_xdp_socket_t **socket, const char *if_name, int if_queue,
 		return ret;
 	}
 
+	(*socket)->frame_limit = FRAME_SIZE;
+	ret = knot_eth_mtu(if_name);
+	if (ret > 0) {
+		(*socket)->frame_limit = MIN((unsigned)ret, (*socket)->frame_limit);
+	}
+
 	ret = kxsk_socket_start(iface, listen_port, (*socket)->xsk);
 	if (ret != KNOT_EOK) {
 		xsk_socket__delete((*socket)->xsk);
@@ -339,7 +346,8 @@ int knot_xdp_send(knot_xdp_socket_t *socket, const knot_xdp_msg_t msgs[],
 			size_t hdr_len = prot_write_hdrs_len(msg);
 			size_t tot_len = hdr_len + msg->payload.iov_len;
 			uint8_t *msg_beg = msg->payload.iov_base - hdr_len;
-			prot_write_eth(msg_beg, msg, msg_beg + tot_len);
+			prot_write_eth(msg_beg, msg, msg_beg + tot_len,
+			               socket->frame_limit - hdr_len);
 
 			*xsk_ring_prod__tx_desc(&socket->tx, idx++) = (struct xdp_desc) {
 				.addr = msg_beg - socket->umem->frames->bytes,