0

I have several physical network interfaces on a debian 6 box, each has a different subnet IP address. Two of the interfaces are connected to a switch hub so they can communicate with each other. I want to setup a TCP connection between the 2 interfaces.

This is what I do: I first get a SOCK_STREAM socket. I then use setsockopt(..., SO_BINDTODEVICE,...) to bind the socket to a specific interface, eth4. I then setup a sockaddr_in structure and populate it with the port and address of that interface (eth4). I then call bind() to bind that local address/port to the socket. I then do a listen() followed by an accept(). That blocks waiting for an incoming connection request. At this point I can do a netstat and it shows exactly what I expect, that is, the socket is indeed in LISTEN state and the Local Address show the eth4 address and port correctly.

Then in a separate program, I have a routine that gets a socket, then setsockopt(..., SO_BINDTODEVICE,...) to bind the socket to a different interface, eth5. I then setup a sockaddr_in structure using the same port number I used for the eth4 interface, and the address of the eth4 interface. I then do a connect(). All goes well, the connect succeeds and my logging indeed shows that the 2 processes are indeed connected with this TCP connection.

But then when I do a netstat, it shows that the 2 PID/Programs as being connected, but both the Local Address and the Foreign address show the same address from eth4 even though I specifically bound the connect end to the address of eth5.

So, my question is why does netstat show the same address for both sides of the connection?

The only explanation I can come up with is that the kernel recognizes that the connect destination address is actually local and then ignores the SO_BINDTODEVICE binding I did and makes the TCP connection internal instead of actually going out eth5 to connect to eth4.

If that is correct, how do I then actually make a TCP connection between 2 separate interfaces, on the same linux box, but actually going external instead of internal?

-Andres

Claudio
  • 10,614
  • 4
  • 31
  • 71
Andres Gonzalez
  • 2,129
  • 5
  • 29
  • 43

1 Answers1

0

You can try calling bind before connect for a client socket to associate it with the particular ip address of the outgoing interface (eth5) you are going to use (port can be set to 0, since it's insignificant in this case for client). This approach seems to be an alternative to using SO_BINDTODEVICE option.

I have found a post offering several ways of setting this option (http://stackoverflow.com/a/6566111/276274). The first one is passing a character string to setsockopt with the device name and another one is passing struct ifreq with the interface index filled in via ioctl. You can try exploring the second one with an interface index.

Another potential source of problems might be the fact that setting this option seems to require root privileges.

It might be helpful, if you could post a code snippet where you are setting the option.

Maksim Skurydzin
  • 10,301
  • 8
  • 40
  • 53