I agree that UDP is connection less. But RFC 1122, which mandates the behaviour of Linux for UDP protocol states that whenever ICMP error messages are received, they should be reported to the user application.

How do you propose this is done? Making the actual send() wait for a possible ICMP package is not practical, as you do not know whether there will come one, and you will not get a success message if the package was accepted.

It is not practical. I am not going to use it. My pursuit is not to know whether the code snippet I wrote is practical, it is - why isn't the application behaving as I think it should? My point is: I should be getting 'Destination Unreachable' errors like I got for 'Connection Refused'.

saurabh.hirani:

You don't understand: You'll receive an error only if the machine sends you an error packet which makes it all the way back to your machine. The 'Connection Refused' is a message from the destination machine¹ telling you to go away. No machine means there's nothing to receive the packet and send you an error result.

Why can't some other computer send you a 'Destination Unreachable' message? Well:

1. Whose responsibility is it to do that? No one has that responsibility.
2. How can they tell that the host doesn't exist? Since a computer doesn't have to support any network services, there's no way for a different computer on the network to tell if the computer actually exists or not.

If you require feedback, then you need to use a connection with TCP.

...roboticus

1. Usually it's the computer. Some firewalls, however, can be configured to send a 'Connection Refused' because the network administrator doesn't want that sort of inbound packet.