SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


    [Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

    RE: Problem with use of NotUnderstood in negotiations



    Julian,
     
    Silent data corruption is when errors change the data in a way that is not detectable. It should be a rare event in a well designed system but it does occasionally occur. For iSCSI, silent data corruption would require:
     
    Error changes the payload such that the link layer CRC check passes (very low probablility - roughly
    (BER^3)*2^32) or the error occurs when the payload is not protected by link layer CRC (e.g. soft memory errors while in switch/router memory).
    AND
    TCP checksum escape (TCP checksum matches the corrupt data)
     
    Once in operational mode, we have the option of using digest to put a further barrier against silent data corruption. We have designed at least some protocol options (digests) for dealing with checksum escape.
     
    I think the approach you suggest for dealing with this is appropriate.
    Shouldn't this condition also be added to the list of errors in 5.10 Negotiation Failures?
     
    For the text in 4.2, it isn't clear what the must applies to. One might use:
     
    An iSCSI device MAY terminate a negotiation which does not end within a reasonable time or number of exchanges.
     
    For 8.3 it looks like MAY and SHOULD should be upper case.
     
    Regards,
    Pat
     
     
    -----Original Message-----
    From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
    Sent: Tuesday, August 13, 2002 12:03 AM
    To: pat_thaler@agilent.com
    Cc: ips@ece.cmu.edu; owner-ips@ece.cmu.edu; wrstuden@wasabisystems.com
    Subject: RE: Problem with use of NotUnderstood in negotiations


    Pat,

    I don't know what "silent data corruption is". The protocol is not built for this type of errors.
    Other than that you observe correctly that we do not mandate the receiver to check for every
    protocol error. However if it does not it has to implement some other means of protecting itself.
    None of the things that where suggested as protection are (or should be) mandatory.
    As for the count/timeout we may want to add to 4.2:

    A negotiation must end in within a reasonable time and/or number of exchanges.

    as well as add to 8.3

    Text negotiations may be also subject to either time-limits or lim-its in the number of exchanges. Those should be generous enough to avoid affecting interoperability (e.g., allowing each key to be nego-tiated on a separate exchange).

    Julo



    pat_thaler@agilent.com
    Sent by: owner-ips@ece.cmu.edu

    08/13/2002 03:35 AM

           
            To:        Julian Satran/Haifa/IBM@IBMIL, wrstuden@wasabisystems.com
            cc:        ips@ece.cmu.edu
            Subject:        RE: Problem with use of NotUnderstood in negotiations

           


    Julian,
     
    Saying that a violation of a rule is a protocol error is not the same as saying that the receiver must detect and react to the error. I don't think your suggested change to the text changes anything. If we want to say that receiving an offer for any key (even if the key is not understood) with a value of "Reject", "Irrelevant" or "NotUnderstood" MUST be detected and treated as a protocol failure, then a specific statement to that effect should be added.
     
    Secondly, there is no requirement that re-negotiation MUST be detected. The specific text on renegotiation is (for login):

    If an attempt to re-negotiate/redeclare parameters not specifically allowed is detected by the target the target MUST respond with Login reject (initiator error); if detected by the initiator the initiator MUST drop the connection.

    The text for operational negotiation is similar except that the actions taken on detection are different.
     
    It says "If an attempt ...is detected" which is entirely different than saying such attempts "MUST be detected".
     
    As Bill as pointed out, detecting re-negotiation of unknown parameters is onerous and I am very opposed to adding such a requirement. It adds unnecessary cost for minimal benefit. For the keys that are supported, on just needs a couple of bits per parameter to keep track of whether the state is unnegotiated, received offer, sent offer or negotiation complete and one may need that state even if one wasn't going to detect re-negotiation. Thus, it is pretty trivial to detect re-negotiation of supported keys.
     
    If someone wants to maliciously tie up a negotiation, they can do so by continuously offering new unknown keys and no working non-malicious implementation will keep offering the same key. One would have to save all
    received unsupported keys and each time an unsupported key was received it would have to be compared to the list. This is time and memory consuming and unnecessary. It isn't required by draft 15 and no such requirement should be added.
     
    A simpler solution to concerns about the silent data corruption problem Bill identified is either
     
    1) add in text requiring that an offer the value equal to NotUnderstood be detected as a protocol error even if the key is not supported/not understood
     
    or
     
    2) rely on timing out the negotiation if it doesn't complete after reasonable time or number of exchanges (definition of reasoanble left to the implementation probably).
     
    Personally, I prefer number 2 because it will catch anything that hangs up a negotiation with one test. I don't know that we can find every corner case. Even if we could, it is better to have one test that digs us out for all corner case errors than to put in many specific tests.
     
    However, number 1 would also be acceptable to me.
     
    Regards,
    Pat
     
    -----Original Message-----
    From:
    Julian Satran [mailto:Julian_Satran@il.ibm.com]
    Sent:
    Saturday, August 10, 2002 9:40 PM
    To:
    Bill Studenmund
    Cc:
    ips@ece.cmu.edu
    Subject:
    RE: Problem with use of NotUnderstood in negotiations


    Bill,


    My only intention in the note was to be brief.  If you viewed that as rude I am really sorry.

    I have a lot of mail to reply to and a limited time.


    As for the crux of the matter - if one of the parties receives a key it has not seen before

    with NotUnderstood as value it MUST act on it as a protocol error. The text is unambiguous on this.

    However we may want to strengthen the rule in 4 (and add another line of text) as follows:


    The constants "None", "Reject", "Irrelevant", and "NotUnderstood" are reserved and MUST ONLY be used as described here. Violation of this rule is a protocol error (in particular the use of "Reject", "Irrele-vant", and "NotUnderstood" as proposed values).


    In the second part of my comment I was just saying that if you  failed to implement the above check you still can't enter a loop

    because the double negotiation rule.


    To get you in a loop in which every party think it answers - then either both are buggy or somebody is altering the messages maliciously.

    The protocol is not designed to handle either.


    Julo


    Bill Studenmund <wrstuden@wasabisystems.com>

    08/11/2002 06:11 AM

           
           To:        Julian Satran/Haifa/IBM@IBMIL

           cc:        <ips@ece.cmu.edu>

           Subject:        RE: Problem with use of NotUnderstood in negotiations


         



    On Sat, 10 Aug 2002, Julian Satran wrote:

    > Bill,
    >
    > Perhaps the text is unabiguos but you just ignored the text that forbids
    > it.

    Julian,

    I must say that the tone above is very unbecoming of the author of a
    protocol spec. In the past, I've often gotten snippy comments from you,
    but written them off to, well, I'm not sure what. But it seems that you
    really don't listen to what I say. You read a message, make an
    interpretation, do not question that interpretation, and then you run with
    it. That's really bad. Besides being quite rude and exceedingly arrogant,
    you'll miss things. And the iSCSI spec will suffer for it. Do you think
    whatever is going on between us worth more than a bad iSCSI spec? I don't.

    Also, if you're going to say that the text is unambiguous, please quote
    said text. That makes the discussion much clearer.

    > The use of Notunderstood is limited to responses. Using it as you suggest
    > is a protocol error.

    Julian, I have to ask, what exactly do you think I'm suggesting? From
    parsing your sentence, I read that you are telling me that I'm suggesting
    using NotUnderstood outside its limited scope, of responses.

    As I understand the situation I described, both parties think they are
    using NotUnderstood as a response. How is using it as a response outside
    its use as a response?

    > A repeated use will also violate the "no renegotiation rule".

    Please be VERY VERY VERY careful when saying that. Have you thought about
    what the statement you just made will imply?

    We are (or at least I started) talking about the case where one side
    THOUGHT it sent key X, but somehow it was key Y that made it to the other
    side. Be it a bug in the code on one side or the other, a PCI bus error in
    transfer, a router glitch, or what. Key Y doesn't exist in the spec, and
    it's in the negotiation stream. OGMarker or DataPDPInOrder would be
    examples.

    The point is, BOTH SIDES THINK THEY ARE RESPONDING to a key they don't
    understand. Reading the spec, it succinctly states if you get a key you
    don't understand, you MUST reply "NotUnderstood".

    Getting back to the "repeated use is a protocol violation," how is each
    side supposed to realize that they are seeing "OGMarker=NotUnderstood" for
    the second time, other than by remembering that it saw OGMarker as a
    not-understood key. i.e. by in addition to replying, "NotUnderstood", each
    side has to remember that it responded NotUnderstood to key foo. That
    seems unwise. I can think of at least one DoS attack if that really is
    what implementations do.

    Getting back to addressing the topic of the thread, what is wrong with
    this text, slightly modified from the I proposed in the first message?

    ***
    Any key not understood by the acceptor may be ignored by the acceptor
    without affecting the basic function. However, unless the value for the
    key was "NotUnderstood", the answer for a key not understood MUST be
    key=NotUnderstood. The value "NotUnderstood" for a key not understood MUST
    be considered a protocol violation.
    ***

    As I said before, my main interest is the spec point out that if for a key
    you don't understand you get the value "NotUnderstood" (i.e. the other
    side is telling you it didn't understand a key that it turns out you also
    don't understand), you don't just answer "NotUnderstood". Either
    saying nothing, or considering it a protocol violation (since if we both
    didn't understand the key it should have never gotten into the
    negotiations) are both fine options. I now favor protocol violation as if
    neither side understood the key, it should not be there. If it's there,

    something is wrong.

    Take care,

    Bill






Home

Last updated: Tue Aug 13 19:18:56 2002
11632 messages in chronological order