X3T9.2/91-157 Rev 1
SCSI-3 Queuing Working Group Meeting Minutes
September 16, 1991  1:00 pm -- 5:00 pm
St. Paul, MN

George Penokie chaired this meeting with the following people present:

             Name              Status          Organization
------------------------------ ------ ------------------------------
Mr. Scott Smyers                 P    Apple Computer
Mr. Paul Wolf                    A    Apple Computer
Mr. Edward A. Gardner            A    Digital Equipment Corp.
Mr. I. Dal Allan                 P    ENDL
Mr. Jeffrey L. Williams          A    Hewlett Packard Co.
Mr. George Penokie               P    IBM Corp.
Mr. Geoff Barton                 P    Iomega Corp.
Mr. Lawrence J. Lamers           P    Maxtor Corp.
Mr. John Lohmeyer                P    NCR Corp.
Mr. James McGrath                P    Quantum Corp.
Mr. Doug Pickford                A    Western Digital

11 People Present

George outlined the agreements from the previous meeting in Boulder:

1.     This work applies to SCSI-3, not SCSI-2.
2.     Doug Hagerman was going to investigate the IPI resource management
       methods.
3.     Several people had promised to generate papers, but none had been
       received to date.

Other issues to be addressed include:

1.     Preprocessing requirements prior to enqueing an I/O process.
2.     Starvation prevention -- can initiators be guaranteed a certain amount
       of the target's resources.

Scott Symers presented a proposal that a message be created to say that the
initiator has just filled his last CDB location.  If the initiator insists on
trying to send another CDB, it would then receive QUEUE FULL status.  Scott's
main reason for this proposal is to prevent retries on CDBs due to QUEUE FULL
status.

Jim McGrath presented a handwritten proposal which he plans to submit in paper
format.  Jim wanted a Central Concept:  Separate diskside actions from those
visible to the host (i.e., SCSI bus sequences).

Jim defined four levels of implementation:

1) Normal          cannot accept another command until status is returned for
                   the previous command.

2) FIFO            rule: cannot transfer data for a subsequent command until
                   all previous commands have returned status.  (Jim contends
                   that this is the SCSI-2 model -- others disagreed.)  The
                   necessary queue depth for this implementation is
                   approximately equal to 2.

3) Streamed        No status phase for a command until all previously
                   initiated commands have been completed.  Harder to
                   implement, but still a driver issue.  Reordering allowed. 
                   The necessary queue depth for this implementation is
                   approximately 4 or more.

4) Reordered       Any phase in any order.  Can affect the OS kernal. 
                   Reordering allowed -- The queue depth for this
                   implementation should be large.

Jim identified some problems that need to be dealt with as I/O Processes are 
enqueued: 

       1) QUEUE FULL status
       2) Reservation (RESERVATION CONFLICT status)
       3) Hardware errors (e.g., parity error)

Ed Gardner presented foils with several questions concerning queuing:

[Items 1) and 2) were not presented since Jim McGrath already addressed them.]

3) Must unordered commands (simple tag) to a sequential-access device (tape)
be executed in order?

4) Consider unordered read and write commands to the same disk blocks:

       read (old data), write
       read (new data), write (new data)  <-- what if the write command fails?
       write, read (new data)

He questioned whether the second case should be allowed.

There was discussion on what issues future meetings should and should not
cover.  There was agreement that there should be no discussion on making
changes to the current SCSI-2 queuing model.

The group concluded that the first order of business is to define a queuing
model for SCSI-3 and that the SCSI-3 model need not be SCSI-2 compatible.  The
model should be defined from the Initiator's point of view rather than
describing the internal workings of the Target.

After the model is complete then the necessary additions and deletions to
SCSI-3 to implement the model will be discussed.
 
It was agreed that further work is necessary and two additional SCSI-3 Queuing
Working meetings were scheduled for October and November.  (Note: The November
meeting was later changed from Monday afternoon to Tuesday evening.)  The
meetings are:

       Monday                   Tuesday
       October 14, 1991         November 12, 1991
       6:00 pm -- 9:00 pm       6:00 pm -- 9:00 pm
       Ft. Lauderdale, FL       Albuquerque, NM