|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: iSCSI : Holes in StatSNJulian, While I agree data delivery directed by the target with significant freedom is the convention used by SCSI, Santosh also speaks to a much larger issue. The present iSCSI transport is not reliable and repairing this weakness using the SCSI layer is a poor technique. Abandon a mixture of SCSI and iSCSI and make iSCSI a self sustaining independent layer. View iSCSI as a general transport that allows handling of digest errors, validation, multiple connections, and sequential delivery then its header may look something like this: +---------------+---------------+---------------+---------------+ 0| Type | Reserved | Word Length | +---------------+---------------+---------------+---------------+ 4| Validation Tag | +---------------+---------------+---------------+---------------+ 8| Checksum (Adler32) | +---------------+---------------+---------------+---------------+ 12| ->ClientSN or <-ServerSN | +---------------+---------------+---------------+---------------+ 16| ->ExpServerSN or <-ExpClientSN | +---------------+---------------+---------------+---------------+ 20| ->MaxServerSN or <-MaxClientSN | +---------------+---------------+---------------+---------------+ +---------------+---------------+---------------+---------------+ 24| (Frame Check) | +---------------+---------------+---------------+---------------+ 28| SCSI_CMND ... | +- | 36| | +---------------+---------------+---------------+---------------+ 40| CDB ... | +---------------+---------------+---------------+---------------+ x | (Data ...) | +---------------+---------------+---------------+---------------+ The advantage of moving the Frame Check (if used) into the header would be with respect to data placement especially in regard to potential overlays or non-sequential data due to the freedom offered to SCSI. To assist in delineating this transport as independent of SCSI a better name could be SCSI Encapsulation Transport or SET. To indicate its use under IP, iSET. The frame type could include FC or Parallel versions of SCSI to assist in a more straight forward conversion. Doug > Santosh, > > Overlaps and out of order delivery and gaps are not forbidden by SAM . I > think we have to go to T10 for that I can't see a good reason to do it. We > have a good solution without asking for it . I can see large and > important > future application that will relay on overlaps and/or gaps and I am not > going to foolishly do something to disable or harm their efficient > implementation. I think that T10s philosophy of keeping the target master > of the transfer and not limiting it in any way is too valuable to ignore. > > IMHO your request violates our charter without any good reason to support > it. > > Julo > > Santosh Rao <santoshr@cup.hp.com> on 01/02/2001 03:53:13 > > Please respond to Santosh Rao <santoshr@cup.hp.com> > > To: ips@ece.cmu.edu > cc: > Subject: Re: iSCSI : Holes in StatSN > > > > > julian_satran@il.ibm.com wrote: > > > With a data sequence we may want to use a similar mechanism to ask for a > > missed data block as soon as we see one of its successors or the status. > > Julian, > > The missing data PDU is missing due to either a header format > error, header > digest error or data digest error. [in all other cases, TCP ensures > reliable > delivery]. > > In 2 of the above 3 cases, [header format error & header digest error] the > initiator CANNOT do a safe interpretation of the PDU header. Without > interpreting the PDU header, the initiator does NOT get the Initiator Task > Tag. Any request to re-send a particular data PDU MUST be qualified by : > I.T.T + missing_DataSN [+ T.T.T + CmdSN, optionally]. > > Since I.T.T. cannot be reliably determined in 2 of the 3 cases, such a > re-send request cannot be reliably achieved. > > The alternate proposal that was made should be considered in its place, > which > was to : > - dis-allow overlapped data xfer's > - initiators do a count check > - a command level retry is performed at the iSCSI layer on detecting an > underrun [due to a missing PDU]. > > On several ocassions, requests from different people have been > made on this > list to dis-allow overlapped data xfers. Can a WG consensus be sought on > this > issue to see if the benefits of allowing overlapped data xfer's offset its > complexities and justify its support ? > > Regards, > Santosh > > - santoshr.vcf > > > >
Home Last updated: Tue Sep 04 01:05:36 2001 6315 messages in chronological order |