SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


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

    Re: iSCSI: definition of TSID (2.11.4 of draft 07)



    
    Mallikarjun,
    
    Comments in line.
    
    Jim Hafner
    
    
    "Mallikarjun C." <cbm@rose.hp.com>@ece.cmu.edu on 08/30/2001 09:57:09 pm
    
    Please respond to cbm@rose.hp.com
    
    Sent by:  owner-ips@ece.cmu.edu
    
    
    To:   ips@ece.cmu.edu
    cc:
    Subject:  Re: iSCSI: definition of TSID (2.11.4 of draft 07)
    
    
    
    Jim,
    
    Sorry, I was travelling, so couldn't get back sooner.  Let me say first
    off that I am not disagreeing with you, only suggesting more explicit
    wording.
    
    I see that the text you sent offline does contain "assumption A" stated
    explicitly - thanks.  I believe that's necessary since one could very
    well
    attach new connections to an existing SSID when its uniqueness is
    maintained
    between just the target portal group and the initiator.
    
    I am still though missing the rationale for mandating uniqueness of SSID
    between the target (across all portals) and initiator.  I assume target
    knows the portal group based on which IP address the login arrives on -
    eventhough not explicitly carried in the payload.
    
    <JLH>
    Are you saying that iSCSI *doesn't* need to ensure uniquess of SSID between
    a single iSCSI initiator and single iSCSI target?  Is that because of
    "locality" within the portal group?  Curious!   I had asked a number of
    months ago what the value of the SSID was in the first place!  I had
    thought that it was to parallel the S_ID, D_ID of FC, but the answer I got
    was "NO", it's just to give each end an inband handle for a session
    context.  I took that to imply some uniqueness.  Then I leveraged that for
    the model.
    
    But I think there might be problems (for one end or the other) if SSIDs are
    not unique.  Suppose initiator A opens one session (ISID=1) to target B on
    Portal Group X and gets TSID=1.  Then A opens session (ISID=1) to B on TG Y
    and again gets TSID=1.  Now, B has done its job of enforcing no repeated
    ISID *to the same portal group* but A may be confused when a messages comes
    from B with SSID=(1,1).  Unless it *also* tracks the TCP connection as part
    of the "session context", it can't tell to what session the message
    belongs.  [Note that in this example, B didn't 'partition' TSIDs between
    its portal groups -- this is what enabled the duplicate SSID to be
    created.]
    
    In short, it is possible for A to "live with" multiple independent sessions
    with the same SSID, but it crosses layering (it requires non-iSCSI protocol
    handles).  In a nut shell, I think we should require this SSID uniqueness.
    </JLH>
    
    Coming to my harping on the nexus identifier, there are two reasons -
     -  It seems to provide a nice rationale for the above issue (refer
        to my previous mail at the bottom on how).  Please set me straight
        if I am totally ignoring something I shouldn't.
     -  You stated that iSCSI is the first transport which opens the
        possibility of "parallel nexus" and so iSCSI mandates implementations
        not to allow it - but appears to stop short of specifying an
        identifier/means to enforce this mandatory requirement.  May be
        it's just me who is troubled by this...
    <JLH>
    Well, yes, we haven't specified *directly* an identifier/means for
    enforcing no parallel nexus.  We have specified the ISID RULE.  Remember
    that I said that the nexus is a "relationship" (not an identifier) between
    two specific SCSI ports; the ISID RULE prevents two independent such
    relationships.  We don't have an explicit (SCSI) nexus identifier mechanism
    for checking, but we have an explicit iSCSI identifier mechanism for
    checking!.  But, if you prevent the relationship from coming into
    existence, then you get no matching nexus identifiers (regardless of how
    that might be reasonably defined) for free.
    
    Is this another one of those places where we're not explicit enough in the
    text?  :-{)
    
    To ease your mind, I have no objection to adding a sentence that defines
    the nexus identifier as the conjunction of the two SCSI portnames and
    observe that by this definition and the ISID RULE, two nexus with the same
    identifier should never occur.
    Does that cover it?
    </JLH>
    
    
    
    
    Regards.
    --
    Mallikarjun
    
    Mallikarjun Chadalapaka
    Networked Storage Architecture
    Network Storage Solutions Organization
    MS 5668 Hewlett-Packard, Roseville.
    cbm@rose.hp.com
    
    
    Jim Hafner wrote:
    >
    > Mallikarjun,
    >
    > There are two assumptions that go into the model and requirements (as I
    see
    > it).  One is currently implicit in draft-07 and is an iSCSI issue, the
    > other is a bit more explicit and is a SCSI issue with iSCSI implications.
    > These are
    >
    > A) (Uniqueness of SIDs) Between a given iSCSI Initiator Node and iSCSI
    > Target Node there can be only one session with a given SID.  [This is
    > implied by the language that says that a login attempt that uses an
    > existing SID=ISID||TSID is an attempt to *add* a connection to the
    existing
    > session. (This is the response I got from Julian when I suggested making
    > this explicit.)
    >
    > B) (No parallel I_T nexus) Between a given SCSI Initiator Port and SCSI
    > Target Port there can be only one nexus.  [With the definition of SCSI
    > Initiator Port as the endpoint of a session identified by iSCSI Initiator
    > Name + ISID, and with the definition of the SCSI Target Port as the
    target
    > portal group identified by the iSCSI Target Name + portal group tag, we
    get
    > the ISID RULE as stated (or intended) in the draft.]
    >
    > The minimalist requirement is to meet the these two rules (assumptions).
    >
    > There is a typo in the text you quote.  It should read "The TSID name
    space
    > of the iSCSI Target should be partitioned among the target portal
    groups".
    > But as we've discussed this really less of a requirement than it is a
    > desirable "note to implementers" (and I'm going to propose that this sort
    > of text move to that section).
    >
    > In what you say below, the "no parallel nexus" (assumption B) is enforced
    > by the ISID RULE.  Assumption A (uniquely identified sessions) is
    enforced
    > if the target does the right thing when selecting a TSID *to pair with*
    an
    > ISID.  My text for the definition of TSID simply says that the target
    > should find a new TSID that enforces Assumption A (no more and no less).
    > All the stuff about partitioning name space belongs in "notes to
    > implementers".  So, between my proposed definition of TSID and the ISID
    > RULE, both assumptions are satisfied and the model works.  No finer
    > restrictions are required *by the model*.
    >
    > Note that in all this description, we haven't had to provide a formal
    > definition of "nexus identifier".
    > The reason is two fold.  First, independent of what we might use to
    > identify them, the I_T nexus is defined as a relationship between a SCSI
    > Initiator Port and a SCSI Target Port.  Assumption B says we can't have
    two
    > such relationships active at one time.  As I've noted above, the ISID
    RULE
    > enforces that.  Second, the nexus identifier doesn't get used in SAM-2
    > explicitly anyway and in SAM-3 it will probably be "up to the transport
    to
    > define AND relative to the viewpoint of the initiator or target".  That
    > last part means that the nexus identifier may not need a globally unique
    > value, but may depend on whose end of the nexus you're looking from.
    E.g.,
    > the target side need only track the "relative target port" and need not
    > care about the external address/name/identifier of that port when
    defining
    > its value for the nexus identifier.  In short, the nexus identifier issue
    > is tangential (and we can safely skip it formally).
    >
    > Jim Hafner
    >
    > [P.S. See also the proposed rewrite of the relevant draft sections in the
    > stuff I sent you off-line, as well, and see if that doesn't help
    clarify.]
    >
    > "Mallikarjun C." <cbm@rose.hp.com>@ece.cmu.edu on 08/28/2001 06:20:16 pm
    >
    > Please respond to cbm@rose.hp.com
    >
    > Sent by:  owner-ips@ece.cmu.edu
    >
    > To:   ips@ece.cmu.edu
    > cc:
    > Subject:  RE: iSCSI: definition of TSID (2.11.4 of draft 07)
    >
    > Jim,
    >
    > Thanks for the explanation.
    >
    > >The model says that the *combination of ISID || TSID* has to be unique
    > >between a given initiator and a given target.
    >
    > Actually, I do not find this formulation stated anywhere
    > in rev07.
    >
    > The closest about TSID assignment policy is this sentence
    > in section 1.5.3: "The TSID name space of the iSCSI Initiator
    > should be partitioned among the target portal groups."
    > My earlier comments were based on this quote.
    >
    > More importantly, let me try to understand the reasoning
    > behind this formulation.  The only reason that I could come
    > up with is to preclude a "parallel nexus".  In our previous
    > conversation, we identified two types of nexus ids for this
    > enforcement - a)InitiatorName-ISID-TargetName-TargetPortalGroupTag
    > b)SSID-InitiatorName-TargetName.  By choosing the stated
    > formulation (in conjunction with the ISID RULE), are you
    > attempting to ensure that a parallel nexus cannot be created
    > if one goes with either definition?
    >
    > Thanks.
    > --
    > Mallikarjun
    >
    > Mallikarjun Chadalapaka
    > Networked Storage Architecture
    > Network Storage Solutions Organization
    > MS 5668   Hewlett-Packard, Roseville.
    > cbm@rose.hp.com
    >
    > >Mallikarjun,
    > >
    > >What you suggest is *more* than is required for uniqueness!  Again,
    that's
    > >an implementation choice that goes beyond the model.
    > >
    > >The model says that the *combination of ISID || TSID* has to be unique
    > >between a given initiator and a given target.
    > >When you say " If there are multiple sessions with an initiator, each
    > would
    > >anyway be assigned a different TSID"  you are assuming that the TSID
    alone
    > >provide the uniqueness (between a pair).   It can, but it's more than is
    > >necessary.
    > >
    > >In short, you're still selecting an implementation.   Here's two other
    > >implementations.  Suppose I have exactly one Target Portal Group.  Then
    I
    > >can meet the uniqueness requirements by (a) enforcing the ISID RULE and
    > (b)
    > >using TSID=1 *for every session*!  This one is "minimalist" with respect
    > to
    > >the model.  (Because there is only one portal group, I can leverage the
    > >uniqueness of the ISID to meet the general uniqueness rule -- it's very
    > >different if I have more portal groups (see note below)) The
    "maximalist"
    > >implementation with respect to the model is (a) enforce the ISID RULE
    and
    > >(b') use a unique TSID for each session I have regardless of initiator
    > >name!
    > >
    > >The spec should say what is required to implement the model.  In my
    mind,
    > >that's the rule that says the combination of ISID and TSID (the SID)
    must
    > >be unique between a pair (initiator and target).   And that's all I'm
    > >asking for in the definition of TSID.
    > >
    > >Jim Hafner
    > >
    > >Note: If I had more than one portal group, I might legitimately see the
    > >same ISID (from the same initiator) come in to each portal group.  In
    that
    > >case, I need to use a different TSID for each such occurrence.  If I
    say,
    > >for example, that each portal group always uses it's portal group tag as
    > >it's TSID (not required, just a choice), then I get uniqueness of SID's
    > >again.  This is again, a minimalist implementation with respect to the
    > >model.   This also gives an example of why the TSID namespace "should"
    be
    > >partitioned between the portal groups, so they can always be sure that
    > they
    > >won't accidentally create an SID that is the same as an existing one.
    As
    > >long as the "half" of the SID one portal group generates is different
    from
    > >the half generated by another portal group, there will be no chance of
    > >collision.  [This is delegation of naming authority!!!]  The same
    > principle
    > >holds on the initiator side in how it generates ISIDs.
    > >
    > >Are we getting closer?
    > >
    > >"Mallikarjun C." <cbm@rose.hp.com>@ece.cmu.edu on 08/28/2001 04:10:24 pm
    > >
    > >Please respond to cbm@rose.hp.com
    > >
    > >Sent by:  owner-ips@ece.cmu.edu
    > >
    > >
    > >To:   Jim Hafner/Almaden/IBM@IBMUS
    > >cc:   ips@ece.cmu.edu
    > >Subject:  RE: iSCSI: definition of TSID (2.11.4 of draft 07)
    > >
    > >
    > >
    > >Jim,
    > >
    > >I see your point about the current wording possibly suggesting
    > >an implementation option.
    > >
    > >Can I suggest the following text:
    > >
    > >  The TSID is a target-assigned tag which together with the
    InitiatorName
    > >  uniquely identifies the session within the target.
    > >
    > >It appears to me that it's all that's required for uniqueness, I
    > >don't think we need to mention ISID here.  If there are multiple
    > >sessions with an initiator, each would anyway be assigned a different
    > >TSID.  What am I missing?
    > >--
    > >Mallikarjun
    > >
    > >
    > >Mallikarjun Chadalapaka
    > >Networked Storage Architecture
    > >Network Storage Solutions Organization
    > >MS 5668   Hewlett-Packard, Roseville.
    > >cbm@rose.hp.com
    > >
    > >
    > >>Marj,
    > >>
    > >>Comments in line (as usual...).  But for the person of few words, I'd
    > >agree
    > >>with just this text (your first sentence only):
    > >>
    > >>"The TSID is the target assigned tag for a session with a specific
    named
    > >>initiator
    > >>that, together with the ISID uniquely identifies a session with that
    > >>initiator."
    > >>
    > >>(I moved the word "name" around because the session is with the
    initiator
    > >>not with the name.)
    > >>
    > >>Jim Hafner
    > >>
    > >>
    > >>"KRUEGER,MARJORIE (HP-Roseville,ex1)" <marjorie_krueger@hp.com>
    > >@ece.cmu.edu
    > >>on 08/28/2001 02:46:45 pm
    > >>
    > >>Sent by:  owner-ips@ece.cmu.edu
    > >>
    > >>
    > >>To:   ips@ece.cmu.edu
    > >>cc:
    > >>Subject:  RE: iSCSI: definition of TSID (2.11.4 of draft 07)
    > >>
    > >>
    > >>
    > >>I'm curious why you've phrased it this way.  I think of it more like;
    > >>
    > >>"The TSID is the target assigned tag for a session with a specific
    > >>initiator
    > >>name that, together with the ISID uniquely identifies a session with
    that
    > >>initiator.
    > >><JLH>
    > >>I don't really see the subtle difference between this text and mine, so
    I
    > >>guess I don't have any problem with it.  At least it doesn't bind the
    > >>definition to anything SCSI'ish.
    > >></JLH>
    > >>
    > >>However, the target MUST enforce uniqueness of the SSID for a given I_T
    > >>combination (in order for login and recovery rules to make sense).
    > >><JLH>
    > >>I think I see what you're getting at here, but I don't think the words
    > >>capture it (it looks like an incorrect use of "I_T nexus" (a SCSI
    thing).
    > >>Do you mean:
    > >>
    > >>  "However, the target MUST enforce the uniqueness of the SSID [aside:
    > >>isn't this 'SID'?] for sessions with a given initiator."
    > >>
    > >>If so, I can go with this text, but I think this is stated in other
    > places
    > >>and need not be here.  But I won't argue if it is also here.
    > >></JLH>
    > >>
    > >>The TSID MAY be a unique tag for a session with a specific initiator
    > >name."
    > >><JLH>
    > >>I don't think this belongs at all.  It is an implementation statement
    and
    > >>so outside the scope.  It's a true statement, but I don't think it
    should
    > >>be made explicit (especially in this part of the document -- I'd have
    > less
    > >>problem with this in the "Notes to Implementers" but even then it is
    > >>suggesting a prefered(?) implementation).
    > >></JLH>
    > >>
    > >>I know the last "rule" is more for convenience than to follow the
    > >"modeling
    > >>rules" (but that's the most straight forward way to implement TSID
    > >>assignment?).  From the target viewpoint, it must first set the context
    > of
    > >>the initiator name, secondly the ISID.
    > >><JLH>
    > >>First, it's not necessarily the most straightforward way. It could
    > >generate
    > >>a unique TSID for every session regardless of initiator name.  That
    gives
    > >>it a twobyte pointer to the session context.  That's a lot more
    efficient
    > >>(perhaps) than having to also have an initiator name in that "pointer".
    > >>So, if you go with your implementation, you have to index on the
    > >name+TSID,
    > >>and that's too big, so you need an indirect pointer to the context and
    at
    > >>that point it doesn't really matter what the TSID is relative to other
    > >>TSIDs; you can use the name + joint SID=ISID||TSID as your index (it's
    > >only
    > >>2 bytes more out of a possible 258!).
    > >>
    > >>So, I argue for no statement about implementation options, leave that
    to
    > >>the clever programmers!
    > >></JLH>
    > >>
    > >>Marjorie Krueger
    > >>Networked Storage Architecture
    > >>Networked Storage Solutions Org.
    > >>Hewlett-Packard
    > >>tel: +1 916 785 2656
    > >>fax: +1 916 785 0391
    > >>email: marjorie_krueger@hp.com
    > >>
    > >>> -----Original Message-----
    > >>> From: Jim Hafner [mailto:hafner@almaden.ibm.com]
    > >>> Sent: Tuesday, August 28, 2001 12:59 PM
    > >>> To: ips@ece.cmu.edu
    > >>> Subject: iSCSI: definition of TSID (2.11.4 of draft 07)
    > >>>
    > >>>
    > >>> Folks,
    > >>>
    > >>> The following text defines TSID in referenced document (and with the
    > >>> changes proposed by Mallikarjun):
    > >>>
    > >>> "The TSID is a tag that identifies the SCSI initiator port.
    > >>> TSID is set by
    > >>> the target.  It MUST be valid only in the final response."
    > >>>
    > >>> I'm trying to figure out what the TSID has to do with the
    > >>> SCSI initiator
    > >>> port.  As I understand it, they are completely unrelated (at an
    > >>> architecture level).  In implementation may choose to use the
    > >>> TSID as a
    > >>> pointer to the session in a unique way across all sessions in
    > >>> the target or
    > >>> in a unique way across all sessions with a given named iSCSI
    > >>> initiator or
    > >>> it can do any number of other things.  It is, in my opinion,
    > >>> unrelated to
    > >>> the SCSI initiator port.
    > >>>
    > >>> I would propose the following alternative text:
    > >>>
    > >>> "The TSID is a tag set by the target that, together with the ISID,
    > >>> identifies a unique session with the initiator."
    > >>>
    > >>> Comments?
    > >>>
    > >>> Jim Hafner
    > >>>
    > >>
    > >>
    > >>
    > >>
    > >
    > >
    > >
    > >
    > >
    > >
    
    
    
    


Home

Last updated: Tue Sep 04 01:03:48 2001
6315 messages in chronological order