Results of January 11, 1994 SAM working group review - Please review

Charles Monia, SHR3-2/W4, 237-6757, monia@starch.enet.dec.com 24-Jan-1994 1320 monia at starch.enet.dec.com
Mon Jan 24 10:20:14 PST 1994


Date:	January 24, 1994

From:	Charles Monia
	SAM Technical Editor

To:	Members of X3T10

Subject: Results of January 11, 1994 SAM Working Group Review

References: (a) X3T10/94-028R0, Summary of Comments on SAM Revision 12
	    (b) X3T9.2/86-109 rev 10h, Draft proposed AMERICAN NATIONAL
	        STANDARD X3.131-199x

Enclosures (a) Compilation of items cited from X3T10/94-028R0

On January 11, 1994, the general SCSI-3 Working Group discussed
the review comments received for  SAM, revision 12  along with
the proposed responses prepared by the technical editor. This
memo summarizes the status of each issue. The citations
accompanying an item identify  pertinent  comments or responses
|from reference (a).

Comments and corrections may be submitted via the SCSI reflector
or directly to me via mail, email or FAX as follows:



	Mail:	Digital Equipment Corporation
		334 South Street
		SHR 3-2/W4
		Shrewsbury, MA 01545
	EMAIL:	monia at starch.enet.dec.com (internet)
	Voice:	508-841-6757
	FAX:	508-841-6100




Item 1:	Policy for target return of autosense data.

The technical editor requested clarification on whether autosense data
should be returned Only on the occurrence of an ACA condition or
Whenever sense data is available -- that is on the occurrence of an ACA
condition and at  other times as specified in the SCSI-2 standard (see
reference b, section 7.2.14, page 127)

See:	Response to Sun comment T 014 on page 62.

Status:	Open.

The technical editor will post this item for review on the SCSI
reflector.

Item 2:	Annex A,  Task Set Boundaries.

Although this annex is identified an "informative", its
inclusion  is potentially confusing and may lead to
non-compliant implementations.

See;	(a) Sun comment T 066, page 80;
	(b) Panasonic comment R2 on page 40.



Status:	Annex A to be deleted.

The working group felt that the annex should be removed,. Since
all interested parties were not present, the matter was referred
to the plenary for discussion. Subsequently, the plenary voted
to delete the annex.

As a related issue,  Bob Snively of Sun requested that the
technical editor include table 1 from the Annex C (queuing model
working paper) in the main body of the document. The editor
agreed to the inclusion of the technical information included in
that table.

Item 2: Limitation on maximum length of vendor-specific command
	descriptor blocks.

In SCSI-2, there is no upper limit on the length of a
vendor-specific command descriptor block. This is in conflict
with the 16-byte upper limit for SCSI-3. 

See: Seagate comment 52 on page 50.

Status: No change to SAM.

The working group consensus was that the limitation in CDB
length was acceptable for SCSI-3. 

Item 4: Bit-significance within the  SCSI status byte.

Since new SCSI-3 status values no longer observe  the bit
conventions used in SCSI-2 there may be a compatibility problem
with the installed base.

See: Seagate comment 55 on page 51.

Status: No change to SAM.

The working group agreed  that removal of bit significance
reflected a  decision of the plenary and should, therefore, not
be changed.

Item 5:	Detection of duplicate task identifiers.

SAM states that the requirement for a target to detect a
duplicate task identifier is established by each protocol
specification. Duplicate tags should always be detected by a
target, regardless of the protocol.

See:	Western Digital comment  on page 92 of a), re: SAM page 50,
     	second paragraph,  section 6.5.3.

Status: No change to SAM.

The working group concurred with Jeff Williams' (HP) view that
the overhead for duplicate tag detection would be excessive for
protocols implementing the maximum tag length (64 bits).

Item 6: SCSI-3 support for tagged queuing.

SAM incorrectly implies that queuing is optional in SCSI-3.

See: 	Hewlett-Packard comment 031 on page 9.

Status:	SAM  will be modified to state explicitly that queuing support
	is optional.

The working Group confirmed that queuing is optional in SCSI-3.

Item 7:	Set of task management functions to be implemented when queuing
	is not supported.

SAM fails to state that certain task management functions, such
as CLEAR TASK SET, are optional when the logical unit does not
implement queuing.

See:	Sun comment T 042 on page 70.

Status:	Comment accepted.

SAM will be modified to state that support for the task
management functions referenced in the comment is only required
if the target supports tagged queuing.

Item 8:	Service Delivery  Transaction Ordering.

For a given sender and receiver, SAM requires the service
delivery subsystem to deliver  a series of transactions in the
order in which they were sent. This requirement should be
deleted since some implementations are incapable of preserving
such order.

See:	a) Hewlett Packard comment T 014 on page 4,
	b) Panasonic comment R3 on page 40,
	c) Sun comment T 007 on page 59,
	d) Sun comment T 048 on page 74,
	e) Sun comment T 062 on page 79.

Status:	Comment accepted.

The working group felt that, while some protocol implementations
may preserve the order of transactions, the only way for an
application to guarantee ordering across all interconnects is to
not use queuing when the sequence of requests must be maintained.

With one or two exceptions the working group supported the
deletion of this requirement. The matter was subsequently put
before the plenary with a similar result.

Item 9:	Number of pending task management requests per I_T_L

The implementor's note in section 4.7.3.1, page 39  of SAM,
advises  against a protocol allowing multiple pending task
management requests per I_T_L. This should be changed to a
requirement prohibiting more than one pending task management
request per I_T_L.

See:  	Western Digital comment on page 86 of (a) re: SAM page 35,
	section 4.7.3.1.

Status:	No change to SAM

The working group felt that  imposing the above restriction
would unneccesarily constrain an implementation.

Item 10:	CLEAR ACA while an ACA task is active.

SAM now specifies that a CLEAR ACA issued while an ACA is in
effect and a task is in the Enabled state should unconditionally
abort that task.  In this case, a task abort should only occur
if the Qerr bit was set.

See:	Sun comment T 40 on page 70.

Status:	Open.

The working group's instructions were to: 

1) Obtain input from those not  able to attend the working
group. The technical editor  will subsequently prepare a
proposal and solicit feedback via the reflector for resolution
at the March meeting.

2) Revise SAM to specify that, if an ACA occurs for a command
issued with the ACA flag set to 0, the ACA will be
unconditionally cleared on receipt off the next command.

Item 11:	Protocol specific arguments in remote procedure calls and
		service calls.

SAM makes no allowance for the inclusion of protocol-specific 
parameters within the procedure calls specified.

See:	Hewlett Packard comment 17 on page 6,
	Sun comment T 055 on page 77.

Status:	SAM will be modified

The draft will state that an implementation interface may
include protocol-specific arguments not  described in SAM.

Item 12:	Multiple target indentifiers

SAM allows a target device to have more than one identifier but
does not address the consequences of such an implementation.

See:	a) Sun comments T 008 and T 009 on page 80,
	b) Western Digital comment on page 85, RE: SAM page 33, object
	   definition 4

Status: No change to SAM

In the opinion of the working group, it is the responsibility of
each protocol standard to fully define the behavior associated
with multiple target identifiers.




The following extracts from X3T10/94-028R0 include both the
comment and the original proposed reply (preceded by a right
arrow).



Item 1:

Begin Sun comment:

-----------------------------------------------------------------
T 014	Page 39, Section 6 and page 74, Section 9.1



	The Autosense Data, as an output argument, is not defined
	elsewhere.  In particular, it is missing from the definition
	of section 9.1, the Confirmation returned to the Application
	Client.  The referenced clause should be defined.

>
> The autosense return flag indicates whether or not sense data was
> returned to the autosense buffer. Input from the working group is
> needed to determine the conditions under which autosense
> data is to be returned. ie. Can autosense data be returned for any command
> or only those which complete with a status of CHECK CONDITION or
> COMMAND TERMINATED?
>
> If the latter, then this flag is unnecessary.
-----------------------------------------------------------------
End Sun Comment and response

Item 2

Begin Sin comment
-----------------------------------------------------------------
T 066	Page 78-81, Annex A

	This section should be removed from the standard.  It offers
	nothing to the definitions of the standard.

	It does, however, point out that there is some information
	missing in the definition of task set.  I suggest that the
	information from Table 1 of annex C be included in the
	normative document, probably in section 4.7.3.2 or section 7.

>

> While the information in the body of the standard must agree with the
=> technical requirements set forth in the queuing model, the editor must
> have the latitude to select the clearest method of presentation.
>
> In that regard, I consider table 1 as illustrative material that
> supplements normative behavior described elsewhere.
>
> In my opinion, therefore, inclusion of table 1 is subject to editorial
> discretion.

-----------------------------------------------------------------
End Sun comments

Extracted from Panasonic comments
-----------------------------------------------------------------

R1.  Though the SAM document has been through several major
revisions there is still significant work needed for the document valuable to
the industry.  This is particularly the case with Annexes A and C which 
represent
a significant amount of committee effort but are not consistent with the
remainder of the document. Concepts like queuing and terminology such as
"execute", "task", "response", "confirmation"  are not consistent. The
document will confuse readers in its present state.

> SAM must, of course, be in full technical agreement with the
queuing

> model passed by the committee. Any discrepancies or inconsistencies between
> the body of the document and the queuing model will be corrected. In that
> regard, please cite specific instances where the draft is either
> unclear or at variance with the queuing model.
>
> It is my understanding that once such corrections are made, annex C will be
> deleted. The committees' intent regarding the alternate task set descriptions
> in annex A, however, is not clear to me. I propose that the issue of
> whether or not to retain annex A be resolved in the working group."

R2.  The document requires the use of "Per Logical Unit Task Set
Boundaries" but discusses and provides for other
implementations. This is very confusing. My experience is that
these options in a standard will become requirements in the 
marketplace and therefore should be better documented in the
standard. If the  intent is to provide extensibility through
these options it should be clearly stated.

>

> I believe the discussion of other alternatives in annex A was in
> accordance with the committees' wishes. Please see the previous response.
>
-----------------------------------------------------------------
End Panasonic comments



Item 3
Begin Seagate comment
-----------------------------------------------------------------
52) Should the Group 6 and 7 vendor specific commands now be
limited to 16 or  less bytes?

>
> I believe that issue should be discussed in the working group. Assuming
> the committee concurs with removing the CDB format defintions from SAM,
>
> I assume this would then become an issue to be addressed by the command
> standards.

-----------------------------------------------------------------
End Seagate comment and response



Item 4:

Begin Seagate Comment
-----------------------------------------------------------------
55) I remain concerned that there may be a conflict with the
installed base  which had a presumption of bit significance and the use of the
BUSY bit in the  Task Set Full Status code.

>
> Please feel free to raise the issue in the working group.
>
-----------------------------------------------------------------
End Seagate Comment And Response



Item 5:

Begin Western Digital comment
-----------------------------------------------------------------
pg 50, 6.5.3, 2nd pgf.: I don't understand the first sentence.
Why not always detect overlaps? Give a reason in a note or delete the
protocol specific provision and make it global.

>
> The reason for making it a protocol-specified requirement is due to
> the large tag space for protocols like FCP [I should have
> added "and hence the large perceived overhead to search for duplicate
> identifiers". In any event, I believe this is another item for further
> discussion].
>
-----------------------------------------------------------------
End Western Digital Comment And Response



Item 6

Begin Hewlett Packard  comment
-----------------------------------------------------------------
#031 (E)  Page 45, Section 6.3, Para 2

TASK SET FULL states that it is required if tagged tasks are
supported.  Tagged tasks support is required, therefore remove
the statement about it being optional (the first sentence).

>
> As I understand it, device support for tagged tasks is still optional.
>
> If it is not optional, then devices such as tape drives will have to be
> modified for SCSI-3 compliance. ie. As a minimum, even if such
> devices are limited to one task per logical unit, they will have to
> understand queue tag messages and the like.
>
> I recommend discussing the matter at the next working group.
>
-----------------------------------------------------------------
End Hewlett Packard comment and response

Item 7:

Begin Sun  comment
-----------------------------------------------------------------
T 042	Page 61, Section 7.4

	The examples pertain only to those LUN's having a queue
	management

[Editor's note: The examples  show various task set management
scenarios]

	capability.  It is also possible that typical tape drives,
	sequential data delivery devices (voice recorders, etc.),
	mechanical devices (like media changers), and certain other devices may
	not allow tagged queuing at all.  In this case, the SAM
	should absolutely not force the devices to perform queue
	management, but rather should allow the synchronous delivery of the 
tasks
	under control of the initiator's application client.  

	This is allowed almost everywhere, but an example should be
	included in section 7 to demonstrate that it is allowed.

	In particular, such devices have a slightly different behavior
	with respect to overfilling their very restricted task set.  A
	second task will be treated as an overlapped command, rather
	than a Task Set Full condition.  This must be made clear in
	sections 6.3 and 6.5.3.
>
> Agreed.
>
-----------------------------------------------------------------
End Sun comment and response

Item 8:

Begin Hewlett Packard comment
-----------------------------------------------------------------

#014 (T)  Page 31, Section 4.6, Para 1

You cannot require that the order be preserved in all cases for
a given pair of devices.  For example, if I run in fibre channel
over a fabric and send two command frames, the order is only
guaranteed if I request in-order transmission over the fabric.
I may not want to do this for performance reasons.  I think that
you need to say that order may be "imposable" in the cases
where ordered tasks are sent or some other ordering is required,
but you cannot require it in all cases.

>
> I believe that, although some physical transports, such as Fibre
> Channel, may reorder data in transit, the ordering specified by the
> sender is restored before the transaction is presented to the consumer.
>
> In any event, ordering is implicit in SCSI-2 today. Therefore
> placing a new ordering burden on the application client (e.g., the
> device driver) may lead to implementations that break existing
> code that would otherwise be portable.

>
> I suggest this issue be left open for discussion
> at the next working group.
>
-----------------------------------------------------------------
End Hewlett Packard comment and response

Begin Panasonic comment
-----------------------------------------------------------------
R3.  I am confused by the requirement in clause 4.6 that all
transactions be received in the order they were sent. My
understanding was that some of the SCSI-3 transports do not
maintain order (i.e. P 1394, Fibre Channel and possibly SSA).
>
> I believe that, although some physical transports, such as Fibre
> Channel, may reorder data in transit, the ordering specified by the
> sender is restored before the transaction is presented to the consumer.
>
> In any event, ordering is implicit in SCSI-2 today. Therefore
> placing a new ordering burden on the application client (e.g., the
> device driver) may lead to implementations that break existing
> code that would otherwise be portable.
>
> I suggest this issue be left open for discussion
> at the next working group.
>
-----------------------------------------------------------------
End Panasonic comment and response


Begin Sun Comment
-----------------------------------------------------------------
T 007	Page 31, Section 4.6


	The last paragraph of section 4.6 indicates that the service
	delivery transactions are received in the order in which they
	are sent for a given pair of source and destination devices.
	Fibre Channel and some other channels allow the proper operation 
	of SCSI with out of order delivery of command information.  This
	restriction should be modified to allow the out of order
	delivery of commands if operating system or channel conventions 
	can guarantee the proper behavior of the SCSI targets.  As an
	example, ordering of groups of commands can be enforced by
	the host adapter function or by management of individual commands
	with respect to the acknowledgment processing of commands
	requiring ordering.
>
> I had assumed that, while data may arrive out of order, the
> receiver would restore ordering before presenting the data to the
> consumer.
>
> I am concerned that relaxing the ordering requirement in the manner
> suggested will lead to implementations that break existing host code which
> depends on the implicit ordering provided by SIP/SPI.
>
> I would like to reopen this issue at the next working group.
>
-----------------------------------------------------------------
End Sun Comment And Response



Begin Sun Comment
-----------------------------------------------------------------
T 048	Page 70, section 8.2

	What should the Device Server do if an Abort Task for a task
	with a certain task identifier overlaps or arrives shortly
	before the task with that task identifier?  There is some
	uncertainty about whether the new task is the one to be
	aborted or is indeed a new task that is not to be aborted.
	This is theoretically possible with certain drivers and out-of-order
	delivery fabrics.



	I would suggest that any task with the specified identifier
	be aborted if it arrives at the device server any time from before 
	the task management function arrives until the service response
	from the task management function is transmitted by the device
	server.  Any identical task after that time should be considered
	a new task.  A recommendation that task identifiers be maintained
	unique for some time period after completion of the task would
	make this requirement more robust.


>

> In my opinion, the behavior specified in SAM should be based on a
> sequential delivery model. An SDS implementation
> may, of course, do what it likes provided the behavior visible to
> application clients, device servers and task managers is sequential.
>
> See the reply to item 007. 
>

-----------------------------------------------------------------
End Sun Comment And Response



Begin Sun Comment
-----------------------------------------------------------------
T 062	Page 76, Section 9.2.1

	The last sentence of the section is incorrect as written.
	The initiator should assure that the blocks of data
	are written into the buffer at the correct displacement 
	within the buffer, regardless of the order in which the
	blocks were actually presented to the interface.


> No, the data must be written in the order received. Otherwise, an
> overlapping transfer that's written out of order will result in
> corrupted data.
>
-----------------------------------------------------------------
End Sun Comment And Response



Item 9:

Begin Western Digital comment
-----------------------------------------------------------------
pg 35, 4.7.3.1: "should" should be "shall", shouldn't it?

[Editor's note:

The referenced text states:

" There are no provisions for aborting, canceling or terminating
a task management function. Thus, an implementation should not
allow an initiator to have more than one pending task management
request per logical unit."}

>
> The present wording was added by request to eliminate what was thought
> to be an unnecessary behavioral restriction.
>
> I suggest reopening the issue at the upcoming meeting of X3T10.

-----------------------------------------------------------------
End Western Digital comment and response

Item 10:

Begin Sun comment
-----------------------------------------------------------------
T 040	Page 59, Section 7.3

	The second sentence of the last paragraph may be overly general.
	The CLEAR ACA should only abort tasks if the QErr bit is set to
	one.

> This should be discussed at the next working group. The description
> in the specification reflects inputs from others in the
> working group.
>
-----------------------------------------------------------------
End Sun comment and response



Item 11:

Begin Hewlett Packard comment
-----------------------------------------------------------------
#017 (T)  Page 38, Section 6, General

You also need an optional input and output arguments for
protocol specific information.  For example, FCP returns
more information in status than just sense data and a status
byte.  It also returns residual byte counts.  There is 
no place for this in the Execute Command primitive.

>
> The intent of the specification is to preserve software portability
> through protocol independence. In my opinion, protocol-specific
> parameters are, therefore, inappropriate for SAM interfaces.
>
>
-----------------------------------------------------------------
End Hewlett Packard comment and response

Begin Sun comment
-----------------------------------------------------------------
T 055	Page 73, Section 9.1

	The SCSI Command Request may be excessively restrictive by
	defining a closed-ended list of request parameters.  Is it
	implicit that a protocol private input and output parameter
	list is allowed, or should it be explicitly allowed.

	This is also a valid question for the indication, response,
	and confirmation parameters.

>
> The goal of the architectural model is to specify behavior in a way
> that's protocol independent. Adding protocol- or implementation-specific
> parameters in the manner suggested above will seriously compromise
> that goal and make it impossible to write portable software or
> firmware.
>
> In my opinion, compromising portability will seriouslyjeopardize the
> industry's significant software and firmware investment. Failure to
> protect that investment will discourage migration to newer technologies.
-----------------------------------------------------------------
 End Sun comment and response

Item 12:

Begin Sun comment
-----------------------------------------------------------------
E 008	Page 33, Section 4.7.1

	Shouldn't the Initiator equation be:

		Initiator = 0{Application Client} + 1{Initiator Identifier}1

	If not, a considerable amount of additional information is
	required to indicate what the rules are for the execution of
	tasks across multiple independent ports.

>
> Multiple identifiers should be considered as nothing more than
> an alias for the same physical entity.
>
> I would like input from the working group on this issue. According
> to past feedback, multiple identifiers for the same entity were
> considered acceptable.
>

 009	Page 34, Section 4.7.2

	Shouldn't the Target equation be:

		Target = 1{Logical Unit} + 1{Target Identifier}1

	If not, a considerable amount of additional information is 
	required to indicate what the rulse are for execution of
	tasks to a LUN having multiple target ports within a single
	task.

>
> See reply to comment 008.
>
-----------------------------------------------------------------
End Sun comment and response

Begin Western Digital comment
-----------------------------------------------------------------
	. . . 
Also (more importantly) all of the tools and services provided
by SAM seem to allow for multiple Initiator and Target Identifiers only
on the most simplified level: you can have more than one, but you can't
relate one to the another within the same device. Given this, why don't
we say:

Initiator = 1{Appl Client} + 1{Initiator Identifier}1

Likewise for target. The alternative is adding considerable
complexity to make multiple IDs fully functional.

>
> Multiple I/D's are supposed to be aliases representing the same physical
> device.
> 
> Since this was added by request, I'd like some feedback from the
> working group before I modify the definition. In my opinion, the
> issue of multiple target identifiers is irrelevant to SAM.
> Although the behavioral model requires one unique identifier,
> a system could implement more than one without violating any
> architectural requirements.
>
>
-----------------------------------------------------------------
End Western Digital comment and response





More information about the T10 mailing list