udhcp: bind to device even for ucast packets

There are cases where binding to source IP and
destination IP is insufficient to guarantee sane
xmit netdev.

One case where this can fail is when
route-matching netdev carrier is down (cable
unplugged, wifi disconnected), or the netdev is
admin down. Then all the IP based bindings (bind()
+ connect()) will seemingly succeed but the actual
packet can go out through a default gw path.

Depending on the network this happens on
it can create issues or false alarms. It can
also leak some subnet info across networks that
shouldn't be routed.

As such better be safe than sorry and bind to a
netdev to be sure it's used for xmit.

function                                             old     new   delta
udhcp_send_kernel_packet                             293     336     +43
send_packet                                          182     188      +6
bcast_or_ucast                                        37      43      +6
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 3/0 up/down: 55/0)               Total: 55 bytes

Signed-off-by: Michal Kazior <michal@plume.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>

1 files changed, 17 insertions, 1 deletions

diff --git a/networking/udhcp/packet.c b/networking/udhcp/packet.c
index 51374646d..4d8e005d4 100644
--- a/networking/udhcp/packet.c
+++ b/networking/udhcp/packet.c
@@ -189,7 +189,8 @@ int FAST_FUNC udhcp_send_raw_packet(struct dhcp_packet *dhcp_pkt,
 /* Let the kernel do all the work for packet generation */
 int FAST_FUNC udhcp_send_kernel_packet(struct dhcp_packet *dhcp_pkt,
 		uint32_t source_nip, int source_port,
-		uint32_t dest_nip, int dest_port)
+		uint32_t dest_nip, int dest_port,
+		const char *ifname)
 {
 	struct sockaddr_in sa;
 	unsigned padding;
@@ -204,6 +205,21 @@ int FAST_FUNC udhcp_send_kernel_packet(struct dhcp_packet *dhcp_pkt,
 	}
 	setsockopt_reuseaddr(fd);
 
+	/* If interface carrier goes down, unless we
+	 * bind socket to a particular netdev, the packet
+	 * can go out through another interface, eg. via
+	 * default route despite being bound to a specific
+	 * source IP. As such, bind to device hard and fail
+	 * otherwise. Sending renewal packets on foreign
+	 * interfaces makes no sense.
+	 */
+	if (ifname) {
+		if (setsockopt_bindtodevice(fd, ifname) < 0) {
+			msg = "bindtodevice";
+			goto ret_close;
+		}
+	}
+
 	memset(&sa, 0, sizeof(sa));
 	sa.sin_family = AF_INET;
 	sa.sin_port = htons(source_port);