SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


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

    Re: iSCSI: DataRN



    Julian,
    
    The current draft says:
    
    1.2.2.2 Response/Status numbering
    
    Responses in transit from the target to the initiator are numbered.
    ...
    Initiators and Targets MUST support the response-numbering scheme
    regardless of the support for command recovery.
    
    The words "are numbered" and "MUST support" don't sound like "should" to me...
    
    -Matt
    
    julian_satran@il.ibm.com wrote:
    > 
    > Santosh,
    > 
    > The current draft has status numbering as SHOULD and that is (as it was
    > pointed out to me) almost mandatory.
    > Considering that the consensus at Orlando and on the list was to make CmdRN
    > (now CmdSN) mandatory even for one connection I don't see any reason not to
    > make StatSN support also mandatory .
    > 
    > On Writes our assumptions was that the target can at any point in time
    > either reject the restart (if as you outline it can't do it) or if it can
    > reacquire the data it misses trough R2T and resend the status (on writes
    > the target is assumed to know what it needs). In case it rejects the
    > restart you can report a "transport failure" and you are as good as a
    > parallel packetized SCSI.
    > 
    > Julo
    > 
    > Santosh Rao <santoshr@cup.hp.com> on 25/01/2001 00:13:53
    > 
    > Please respond to Santosh Rao <santoshr@cup.hp.com>
    > 
    > To:   Julian Satran/Haifa/IBM@IBMIL
    > cc:   ips@ece.cmu.edu
    > Subject:  Re: iSCSI: DataRN
    > 
    > julian_satran@il.ibm.com wrote:
    > 
    > > Santosh,
    > >
    > > The ACK for a response is when status is ACKed (by the status numbering).
    > >
    > > Before this ack a target has several options for restart:
    > >
    > >    to keep all the data (recall we have dropped DataRN) and "restart
    > >    sending".
    > 
    > This may work for a READ if the iSCSI layer has buffered the data until it
    > gets a StatSN ACK. Note that even for the READ, status recovery is optional
    > and a tape may not implement status / data recovery and just retry the
    > command. In such a case, a retry results in the target re-doing the
    > operation.
    > 
    > For a WRITE however, I don't see how this is going to work. Are you
    > suggesting
    > that the iSCSI layer NOT deliver data to the SCSI layer until the Response
    > is
    > acknowleged ??? This is ruled out, since the Response itself comes from the
    > SCSI layer upon completion of the SCSI command.
    > 
    > If, on the WRITE, the data was delivered to the SCSI layer which then wrote
    > it
    > to media and advanced the tape, what good does the above do ?? The tape has
    > advanced following the WRITE and the initiator can corrupt the tape by
    > retrying the WRITE.
    > 
    > >    not to keep data and reject the command restart with a service
    > response
    > >    (that we have to specify) of restart reject (your tape target may want
    > >    to do just this)
    > 
    > You seem to suggest moving the responsibility of ensuring data integrity
    > from
    > the host to the target !! This is not suitable for the following reasons :
    > a) Often, tape is the last class of storage products to migrate to new
    > storage
    > transport technologies. IOW, the majority of tape backup is going to attain
    > iSCSI connectivity through a iSCSI-FC bridge or iSCS-pSCSI bridge. In such
    > a
    > case, the service response management is done in the bridge and this level
    > of
    > intelligence is now being moved into the bridge.
    > 
    > The responsiblity of ensuring data integrity on a SCSI target lies with the
    > initiator issuing the I/O.
    > 
    > iSCSI MUST not retry a command to sequential media when there is the
    > possiblity that some or all of the data phase has completed.
    > 
    > Regards,
    > Santosh
    > 
    >  - santoshr.vcf
    


Home

Last updated: Tue Sep 04 01:05:44 2001
6315 messages in chronological order