SCTP multi-homed retransmissions

Discussion in 'Linux Networking' started by bjeremy32, May 14, 2013.

  1. bjeremy32

    bjeremy32 Guest

    Firstly, if this is not the correct group, I apologize it seemed the closest match. My question involves a SCTP multi-homed scenario. It seems a colleague and I are at a little bit of disagreement.

    The scenario begins with retransmitted control chunks before the association is setup. There are two multi-homed endpoints (a1, a2) -> (b1,b2). Due toa routing problem b1 is not reachable from a1, but it is from a2. Also, a1is reachable from b1 and b2.

    The Init is sent out from a1 to b1. This reaches the RTO and then the stackretransmits out from a2 to b2. This reaches b2 and the INIT-ACK is sent out to a2, however the source address it is sent from is b1, not b2 as it hasbeen sent.

    The Cookie-echo chunk is then first sent from a1 back to b1. This times outagain, and our stack is not trying to send a retransmit from a2. It just keeps sending out from a1 until it times out and reaches the max number of re-transmits.

    Now my colleague is trying to say this is expected behavior. Since the INIT-ACK came from b1, and since b1 is unreachable we should fail sending, and that the INIT-ACK should have come from b2. Since the INIT-ACK came from b1, it is perfectly valid to assume b1 is a valid ip and to just keep trying to send to b1 without trying b2. Since b2 did not respond we should consider b2 invalid.

    I whole-heartedly disagree. I say that the specs only say that a reply mustbe destined to the remote peer endpoint's transport address of origination. They do not specify which of the local endpoint's transport addresses itis sent from. And also, if a1 tries to send the cookie-echo chunk to b1 and this times out, we must actually try the secondary transport address b2, similar to how the INIT behaved.

    My colleague cited references in the rfc section 5.1 and 5.1.2, but they only seemed to state how the message is to be routed by destination, they make no mention of from where the message originates. And I further cited section 6.4 which specifies that in a multi-homed scenario if the chunk can notbe sent to the first address we should send to the second.

    Anyway, I was I was wondering what insight others had into interpreting thespecs.
     
    bjeremy32, May 14, 2013
    #1
    1. Advertisements

  2. bjeremy32

    Jorgen Grahn Guest

    comp.protocols.tcp-ip is the right group, although I haven't seen SCTP
    being discussed there recently.

    Try to keep your line lengths down if you repost there.

    /Jorgen
     
    Jorgen Grahn, May 14, 2013
    #2
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.