SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


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

    iSCSI: Yokohama meeting technical issues



    The following iSCSI technical issues were discussed
    at the Yokohama meeting (minutes are coming ...).
    
    The following paragraphs report the rough consensus
    of the room in Yokohama and I believe they are the rough
    consensus of the IP Storage WG.  In Yokohama, some of
    these consensus calls were over the objection of one
    person (not the same person each time).  This is an
    opportunity for those who couldn't make it to Yokohama
    to voice a different opinion:
    
    (1) Long Lived Discovery sessions.  Support for these
    is not intended.  The only PDUs that an initiator is
    allowed to send on a Discovery Session are a Send Targets
    text request and a logout request that specifies "close
    the session".  Targets MUST reject all other PDUs.
    
    (2) Decimal encoding.  Follows the rough consensus on
    the list, decimal encoding can only be used when the
    parameter being encoded has a fixed size (bits that
    the decoded value occupies) and that size is 64 bits
    or less.  If one of the authentication bignums happens
    to have a value that fits in 64 bits, decimal cannot
    be used because the bignum has a considerably larger
    size.
    
    (3) Size requirements for negotiation.  Compromised on
    8kb as the minimum amount of text that MUST be accepted
    during a negotiation (consensus called over one person's
    objection).  64k limit applies only when an AuthMethod is
    supported that requires large binary values - in the meeting,
    this was thought to be only SPKM, but a quick check of
    the draft suggests that the 64k limit also has to apply
    apply to Kerberos - it does not apply to CHAP or SRP.
    
    (4) Error Recovery Level 0.5.  An implementation that does
    not support both of the major features required for
    ErrorRecoveryLevel 1 (within-connection and within-command
    recovery) MUST negotiate ErrorRecoveryLevel 0, as to
    negotiate 1 would promise support for both.  When
    0 is negotiated, an initiator MAY try within-connection
    recovery or within-command recovery, but it MUST be prepared
    for a target to reject the attempts, as the target may be
    strictly operating at ErrorRecoveryLevel 0.  A new value
    of the ErrorRecoveryLevel key will *NOT* be defined to enable 
    partial support of level 1 to be negotiated.
    
    (5) BidiInitialR2T.  This key will be removed, InitialR2T will
    also cover the bidirectional case, in part because this
    corresponds to Fibre Channel's behavior.
    
    (6) Resegmenting Data SNACKS - an intrepid design team came
    up with a compromise solution between the Monday and Tuesday
    sessions (including a trans-pacific conference call).
    
    Summary of the problem situation:
    
    	- SCSI Read or similar command is active on an
    		iSCSI connection.
    	- The initiator reduces the Maximum Data PDU size it
    		is willing to accept.
    	- The initiator discovers that it's missing some
    		of the inbound data and requests retransmission.
    
    iSCSI's data retransmission is based on resending
    the same PDUs that were sent the first time.  That doesn't
    work in this case because they may be too large.  When they're
    made smaller, the count of Data PDUs in the SCSI Response PDU
    may have to be changed.  This should be a rare case.
    
    A rough summary of the compromise solution:
    
    	- Existing Data SNACK MUST NOT resegment.  Targets
    		MUST check this.
    	- A new R-Data SNACK code is defined for the resegmenting
    		case.  It always requests transmission of all
    		unacknowledged data, and of a new SCSI Response
    		since the count of Data PDUs may have changed.
    		The new SCSI Response uses the same StatSN number.
    	- A new SNACK tag is sent in an R-Data SNACK and
    		returned in the new SCSI Response to ensure
    		that the new response with the correct count
    		of Data PDUs is used, as there may be one or
    		more SCSI Responses to this command with the
    		the wrong count in flight when the R-Data
    		SNACK is issued (if you're wondering how
    		multiple SCSI Responses could be in flight,
    		think about Status SNACKs and large delays).
    
    There are a number of additional details that can be found
    in the working draft.
    
    I believe the above six items represent the rough consensus
    of the IP Storage WG, but this is an opportunity for those
    who differ (and could not make it to the Yokohama meeting)
    to object (quickly, please).
    
    Thanks,
    --David
    ---------------------------------------------------
    David L. Black, Senior Technologist
    EMC Corporation, 42 South St., Hopkinton, MA  01748
    +1 (508) 249-6449            FAX: +1 (508) 497-8018
    black_david@emc.com       Mobile: +1 (978) 394-7754
    ---------------------------------------------------
    


Home

Last updated: Tue Jul 30 10:39:11 2002
11481 messages in chronological order