Can you bind() and connect() both ends of a UDP connection

Question

I'm writing a point-to-point message queue system, and it has to be able to operate over UDP. I could arbitrarily pick one side or the other to be the "server" but it doesn't seem quite right since both ends are sending and receiving the same type of data from the other.

Is it possible to bind() and connect() both ends so that they send/receive only from each other? That seems like a nicely symmetric way to do it.

Answer 1

Hello from the distant future which is the year 2018, to the year 2012.

There's, in fact, a reason behind connect()ing an UDP socket in practice (though blessed POSIX and its implementations don't in theory require you to).

An ordinary UDP socket doesn't know anything about its future destinations, so it performs a route lookup each time sendmsg() is called.

However, if connect() is called beforehand with a particular remote receiver's IP and port, the operating system kernel will be able to write down the reference to the route and assign it to the socket, making it significantly faster to send a message if subsequent sendmsg() calls do not specify a receiver (otherwise the previous setting would be ignored), choosing the default one instead.

Look at the lines 1070 through 1171:

if (connected)     rt = (struct rtable *)sk_dst_check(sk, 0);  if (!rt) {     [..skip..]      rt = ip_route_output_flow(net, fl4, sk);      [..skip..] } 

Until Linux kernel 4.18, this feature had been mostly limited to the IPv4 address family only. However, since 4.18-rc4 (and hopefully Linux kernel release 4.18 as well), it's fully functional with IPv6 sockets as well.

It may be a source of a serious performance benefit, though it will heavily depend on the OS you're using. At least, if you're using Linux and don't use the socket for multiple remote handlers, you should give it a try.

Answer 2

UDP is connectionless, so there's little sense for the OS in actually making some sort of connection.

In BSD sockets one can do a connect on a UDP socket, but this basically just sets the default destination address for send (instead giving explicitly to send_to).

Bind on a UDP socket tells the OS for which incoming, local interface address to actually accept packets (all packets to other addresses are dropped), regardless the kind of socket.

Upon receiving you must use recvfrom to identify which source the packet comes from. Note that if you want some sort of authentication, then using just the addresses involved is as insecure as no lock at all. TCP connections can be hijacked and naked UDP literally has IP spoofing written all over its head. You must add some sort of HMAC