P1394 Mail Archive: P1394> Meeting Minutes from Monterey - July

P1394 Mail Archive: P1394> Meeting Minutes from Monterey - July

P1394> Meeting Minutes from Monterey - July

Gregory LeClair (gleclair@best.com)
Mon, 28 Sep 1998 14:36:12 -0700

===============================================

Meeting Attendees

Takashi Isoda Canon
Osamu Hirata Canon
Akihiro Shimura Canon
Lee Farrell Canon Information Systems
Greg LeClair Epson (1394 PWG Chairman)
Fumio Nagasaka Seiko Epson
Atsushi Uchino Seiko Epson
Brian Batchelder Hewlett Packard
Alan Berkema Hewlett Packard
Scott Bonar Hewlett Packard
Eric Clement Hewlett Packard
Laurie Lasslo Hewlett Packard
Greg Shue Hewlett Packard
Kwang Kim Hitachi
Yuji Sasaki Japan Computer Industry
Brian Nagy Kodak
Jerry Thrasher Lexmark
Don Wright Lexmark
John Fuller Microsoft
Frank Zhao Panasonic
Randy Turner Sharp
Bob Morford SIS Microelectronics
Anthony Fung ST Micro Electronics
Peter Groz ST Micro Electronics
Harry Hvostov ST Micro Electronics

Administrivia

Don Wright gave the next PWG meeting details:
August 17-21
Toronto Marriott at Eaton Centre
525 Bay Street
Toronto, Ontario, Canada M5G 2L2
$175CN (about $120 US)
Registration deadline is July 24, 1998

Future PWG weekly schedules will be as follows:
1394 PWG on Monday and Tuesday
MIB meetings on Tuesday evening (if needed)
PWG Plenary meetings on Wednesday morning
IPP on Wednesday afternoon and Thursday morning
SDP on Thursday afternoon
UPD on Friday

Greg LeClair presented the 1394 PWG meeting goals and proposed agenda
topics:
OUI Issues
Transport Command Set
Config ROM Review
Profile Review
Updates

OUI Issues

Greg LeClair asked if we need to have an OUI for the 1394 PWG Profile?

The group did not resolve this issue, but a few attendees suggested that
the IEEE 1212 group should maintain the list of Function Classes. Greg
said that this topic was raised at the previous 1212 meeting, but no
conclusion was reached at that meeting. Further discussions were
deferred.

Error Recovery Model

Isoda-san gave a presentation on his ideas for supporting error
recovery. His main requirements are that the Initiator shall keep the
contents of the data buffers associated with the ORBs in the linked
list, and maintain the correspondence of the data buffers to a Sequence
Identifier. The Target must guarantee that it does not execute any data
or command twice.

Greg Shue asked if the proposal addresses both writing from the
Initiator as well as writing from the Target. Isoda-san suggested that
it does address both, but he is more confident about the Initiator
writing recovery.

Isoda-san claims that we should maintain a Sequence Id for the ORBs. As
part of the proposal, he explained that it is not necessary for every
ORB to have a status notification sent. It is adequate to send status
for a subset of the ORBs. Specifically, if a Sequence Id can be a
maximum of 2n-1, it is necessary to have at least one status sent for
every group of n ORBs that have been sent since the previous status.

He showed that, based on the requeuing of ORBs, the Target can determine
if the Initiator has received the most recently sent status block. He
also said that if a Sequence Id is not used, the Target cannot determine
if the Initiator has received the status block or not.

A few people commented that the idea of a "sliding window" of ORB
sequences was a good one.

During discussion of the proposal, Randy Turner said that he believes we
should keep negotiation of buffer size.

Brian Batchelder asked if anyone could evaluate the benefit (in terms of
performance efficiency) gained by reducing the number of notification
status blocks sent. The only comment was that "it really depends on the
size of the data buffers."

Transport Command Set

Greg Shue referenced his proposal document, PWG 1394 Transport Command
Set Proposal Revision 0c. (This document is available at
ftp://ftp.pwg.org/pub/pwg/p1394/mtg070698/pwgcmd0c.pdf.) He provided a
quick overview of the document and each of the open issues including
those items in the document that are identified as "to be determined."
The major discussion items and any conclusions reached are given below.
(The discussions were not sequential as the paragraphs below might
imply.)

[NOTE: Greg LeClair captured additional issues that were raised. His
list should be included in the minutes.]

Queues

Should the usage of each queue be specified (e.g. Read and Write) or
left arbitrary? Should we allow a vendor to extend the number of queues
within a single Login?

Brian Batchelder suggested that if an implementer would like to extend
the number of queues, it should be accomplished by additional Logins. He
suggests that we should limit a given Login to a (simple) bi-directional
communication "pipe." (The idea of two independent queues was originally
proposed because SBP-2 doesn't inherently support this capability.)

The group agreed to limit the number of queues to 2; one for
Initiator-to-Target transfer and one for Target-to-Initiator transfer.
It is intended that the group will focus on providing the equivalent of
a bi-directional transport.

Should we eliminate the queue (q) bit, and just use the direction (d)
bit only? Is it a reasonable assumption that the direction is equivalent
to the queue id? It was decided that this issue is dependent on error
recovery, and it was deferred.

Notification

There was a long discussion about the problems that can occur if a
status notification (or its acknowledgment) can be lost. There was also
some doubt expressed about the frequency of this condition ever
happening. If it is an extremely unlikely event, perhaps it is better to
simplify the recovery process as much as possible.

Should we set the notify bit to one (1) for every ORB? Do we need to
always notify completion of every ORB? Greg Shue points out that we are
using the unordered model of the SBP-2 specification. A direct
implication of this is that every ORB must have completion notification.
A "well-partitioned," layered SBP-2 driver does not have sufficient
knowledge about the ORBs in the queue when using an unordered model. If
the driver did have this knowledge (by examining the q bit, for
example), it would be considered "munging" (combining) the different
layers.

After looking at the latest SBP-2 specification, this issue is not
(apparently) explicitly addressed. There is no clear statement that says
the Initiator must set the notify bit when using the unordered model.
Greg Shue would like to make an explicit statement about this in the PWG
Profile document. He identified three possible alternatives:
1. Require an Initiator to set the notify bit on every ORB
2. Require the Target to always give completion notification
3. Allow for the possibility of "monolithic" (multi-layered)
SBP-drivers to be implemented without saying anything additional beyond
what is already specified in the SBP-2 specification

After more discussion, the group agreed that the PWG Profile document
would not make any additional, explicit statements on this issue other
than to follow the SBP-2 specification.

Extensibility

To what degree should we provide extensibility? Do we really need to
support a mechanism for vendor-specific transport commands?

As stated above, the group agreed to limit the number of queues to 2.
The number of queues is not extensible.

Shimura-san suggested that vendor-specific commands could be useful as a
method to fix bugs. Most of the other attendees felt that this was not a
good enough reason to include support for vendor-specific commands. It
was suggested that vendor feature differentiation could (and should?) be
achieved at higher levels.

The group agreed that a Command OUI is not necessary (at least for this
version of the Profile.) No vendor-specific commands or parameters will
be supported. Any extensions will only be possible through the PWG
standards group by revision of the document.

Parameters

When can parameters be set and/or changed?

Because the SBP-2 Login is not extensible, Randy Turner suggests that
the PWG Profile should include a Login Response, followed by an exchange
of additional parameters. He also pointed out that both the Target and
the Initiator should know each other's capabilities -- at least to help
resource allocation efficiency. It was suggested that the commands Get
Param and Set Param List should be executed before data transfer,
similar to a "connect" process. This suggestion was later expanded to an
Open command that could establish other items that would be necessary
for managing a connection.

ACTION: Randy volunteered to write up an example that would explain
this concept.

Error Recovery

Should we include a Sequence Id for the ORBs? What is the scope of the
sequence number?

After a fairly brief discussion, the group agreed to use a Sequence Id
for error recovery purposes.

Brian Batchelder noted that the Target should not process more than 2n
ORBs ahead of the last ORB for which the Target received an
acknowledgment of completion status from the Initiator (where 'n' is the
bit width of the Sequence Id.) Otherwise, it is possible that the Target
could lose track of the error recovery information provided by the
Sequence Id.

A long discussion occurred about pre-fetching ORBs and the reliability
of ORB pointers after a bus reset.

Greg Shue noted that the group needs to define a unit attention status
model. However, he suggested that the topic is probably too large to
adequately address at this meeting. It was suggested that the topic be
discussed via e-mail.

A few individuals discussed the processes involved in error recovery,
based on the assumptions of the proposal given by Isoda-san. After the
discussion, John Fuller proposed that the Sequence Id should be eight
bits wide. Nagasaka-san wanted it to be 16 bits wide. After discussing
the need to have the Sequence Id be sufficiently larger than the Maximum
Task Set, the group agreed to 16 bits for the Sequence Id and 14 bits
for the Maximum Task Set size.

What should be the initial value for the Sequence Id? Greg Shue
suggested that the initial value could be inferred based on the first
ORB on each queue after a Login or an abort task. Randy preferred that
some "initial process" be defined to ensure a specific value for the
first Sequence Id. The topic was deferred for discussion via e-mail.

John Fuller noted that the ORB Sequence Ids might have "gaps," because
of the use of Abort Task and Abort Task Set commands. This caused people
to wonder if it is necessary to generate a notification for each ORB to
handle error recovery properly. However, it was decided that Greg Shue
(and others) would spend time after the meeting to examine this question
more closely.

ACTION: Greg will attempt to write up a statement for discussion via
e-mail.

Randy is concerned that without a timeout value (for an ORB completion
notification), there is no way a host can know the lifetime of an ORB.
How long should an Initiator keep an ORB allocated? The group suggested
that if the Initiator wants to free the resources associated with an ORB
before the completion notification has been received, an abort task
command should be issued.

What should the timeouts be for the following items?
abort task
abort task set
logical unit reset
Target reset

The discussion of timeout values was deferred.

Open/Close Connection

Several items were put into a category of "connection meta-data" that
should be communicated and established. The items include:
initial values for Sequence Id numbers
connection type
completion policy
maximum record size

How can a Target close a connection (i.e. unsolicited status)?

This entire topic was deferred for e-mail discussion.

Other Deferred Topics/Issues/Questions

The following items were raised during the discussions, but not directly
resolved:
When should a Target complete the Store Data Command Response?
Does it make sense to use the names "Store Data" and "Fetch Data" for
commands?
Would it be better to just have a single "Transfer" command, with the
direction (d) bit used to indicate which way the data should be
transferred?
Should the q bit be placed in the status block?
Should we keep the Max Data Size Per Fetch ORB parameter?
Should we keep the Max Data Size Per Store ORB parameter?
What should happen to Fetch Data and Store Data commands that are in
or added to the active task set when a data flow direction is disabled?
Maybe we could put "last acknowledged Sequence Id" in each status
block?
Perhaps we should add some attributes and features to the Unit
Directory? This might be a mechanism (for example) to communicate the
maximum buffer size of the Target.

Summary of Commands Discussions

Greg Shue attempted to summarize the decisions reached:
2 queues only
No vendor-unique extensions (for this revision)
no CMD_OUI
no Parameter_OUI
Provide only equivalent bi-directional transport
MSG (packet)
Stream
Error recovery
16 bit increasing Sequence Ids
14 bit size for Max Task Set

Remaining issues and assignments for write-ups:

sequence # [Greg Shue]
seeding, abort task, interactions
notify bits and gaps in sequence
Open/Close CMDs [Randy Turner and Greg Shue]
parameters and encodings
negotiated buffer sizes
mode and completion policy [Bob Morford]
management agent CMD support
How will Target initiate close (unsolicited status)?
How to communicate/clear UNIT ATTENTION?
Timeouts [Alan Berkema]
Is queue implied by direction bit?
What initiates error recovery on Microsoft stack? What steps are
taken? [John Fuller]

Config ROM Review

Because Greg LeClair could not be present the second day of the meeting,
Alan Berkema attempted to lead a discussion of the CSR and Config ROM
document. (This document is available at
ftp://ftp.pwg.org/pub/pwg/p1394/mtg070698/pwg_cfgrom.pdf.) However, it
did not seem that many people read the document. There was very little
discussion.

Imaging Device Profile Review

Alan Berkema decided that the Profile document will continue to contain
Issues and their resolutions. (This document is available at
ftp://ftp.pwg.org/pub/pwg/p1394/mtg070698/pwg_cfgrom.pdf.)

Meeting adjourned.