RE: iscsi : target port definition (private)

["Jim Hafner" <hafner@almaden.ibm.com>]

Santosh,

I'm not sure I quite caught the thrust of your note.  Are you supportive of
the model or questioning?

I get thrown by your comment:
 "If we change the definition of TSID to be unique within the target, then
I
think this is a good idea."

If I read this as "unique independent of initiator", then, as we've
discussed before, that's an implementation choice that goes far beyond what
any part of the model says (however we define SSID).   I'd be VERY opposed
to making this the requirement.    I'll explain the reasons if (a) my
interpretation of your quote is correct and (b) it is a formal
recommendation to this group.  If not, I'll be quiet on this point.  But I
have no objection to *you* implementing that way. :-{)

You say "the TSID can be the same for different initiators".  In fact, it
can be the same for the same initiator provided the ISID from that
initiator is different (so that the resulting SSID is unique)!  So, I could
implement my target by having each target portal group use its target
portal group tag as the value of the TSID for every session regardless of
initiator.   (That's a very simple partition logic!).  By enforcing the
ISID RULE, every session would get a unique SSID (ergo, I've met the TSID
RULE by default).  [Note: I could do the exact same thing on the initiator,
that is, make every initiator portal group use its portal group tag as its
ISID -- I wouldn't be able to use multiple sessions to the same target
portal group. but that's an implementation choice!]

Of course, such a target implementation wouldn't be able to use the TSID as
a handle to some session reference (though it could use the entire SSID),
but the spec doesn't say how one uses (internally) the TSID, just what
properties it has to have externally.

Are we all on the same page yet?

Jim Hafner


"Eddy Quicksall" <eddy_quicksall@ivivity.com> on 09/14/2001 01:23:56 pm

Please respond to <eddy_quicksall@ivivity.com>

To:   "'Santosh Rao'" <santoshr@cup.hp.com>, Jim Hafner/Almaden/IBM@IBMUS
cc:   <ips@ece.cmu.edu>
Subject:  RE: iscsi : target port definition



Currently the TSID can be the same for different initiators (now, in my
case, they'll all be unique).

If we change the definition of TSID to be unique within the target, then I
think this is a good idea.

But, I like the idea of Portal Groups as being a nice way for an
administrator to allocate the Network Portals and the Portal Group
properties.

TSID RULE: The iSCSI Target SHALL NOT select a TSID for a given login
request if the resulting SSID is already in use by an existing
session between that the target and the requesting iSCSI Initiator.
See 8.1.1.

Eddy

-----Original Message-----
From: Santosh Rao [mailto:santoshr@cup.hp.com]
Sent: Friday, September 14, 2001 1:49 PM
To: Jim Hafner
Cc: ips@ece.cmu.edu
Subject: Re: iscsi : target port definition


Jim,

I agree with your below ideas. However, they do appear to conflict with the
iscsi rev 08 definition of the I-T nexus in Section 1.5.2 which reads :

" c) I_T nexus - According to [SAM2], the I_T nexus is a relationship
between
a SCSI Initiator Port and a  SCSI Target Port.  For iSCSI, this
relationship
is a session, defined as a relationship between an iSCSI Initiator's end of
the session  (SCSI Initiator Port) and the iSCSI Target's Portal Group.
The
I_T nexus can  be identified by the conjunction of the SCSI port names;
that
is, the I_T nexus identifier is the tuple (iSCSI
Initiator Name + 'i'+ ISID, iSCSI Target Name + 't'+ Portal Group Tag).

NOTE: The I_T nexus identifier is not equal to the session identifier
(SSID).
"

Per the above definition, the I-T nexus target end point is that target
portal
group, which is not the case when initiators choose to establish I-T nexi
(sessions) to subsets of the target portal group, in order to export
multiple
scsi paths to upper layer wedge drivers.

Further, initiators that establish sessions to a subset of the target
portal
group would not be able to take advantage of initiator specific mode page
implementations on the target. All the I-T nexi (sessions) established by
seperate initiator ports to a given target portal group would always share
the
same mode pages, since the target mode pages would be implemented on a per
target portal group basis.

Do you see any reasons why the definition of a target port should not be
symmetric with the definition of the initiator port ? i.e. (iscsi target
name
+ TSID) = target port. (= both port name & port identifier). This would
more
accurately model the target port to be the end point of the I-T nexus
(session).

Regards,
Santosh


Jim Hafner wrote:

> Santosh,
>
> There are many alternatives here, but I think the simplest is to
establish
> multiple sessions to the one target portal group.  Each session can
connect
> to one, some or all of the target ipaddresses in the portal group, so you
> have a lot of flexibility.  What you see then in the wedge driver is
> multiple "SCSI Initiator Ports" in the host connecting to one SCSI Target
> Port.  That should be sufficient for the multipathing logic.    So, it's
> many to one, not one to many.
>
> Note that according to the model in -08, if there were more than one
target
> portal group, you could see two different things, depending on how you
> implemented your host.  You could have an "implementation" of one host
SCSI
> Initiator Port connecting to multiple SCSI Target Ports if you used the
> same ISID for all those sessions.  Or you could have an implementation of
> multipel SCSI Initiator Ports connecting in arbitrary ways to the
multiple
> SCSI Target Ports if you used a set of ISIDs.  In other words, the reuse
of
> an ISID to a different target portal group implies a one to many setup.
> And you can overlay lots of one-to-many (or one-to-one) sessions as you
> enable different ISIDs.  In other words, the "many" SCSI Initator Ports
are
> based on multiple use of ISIDs and multiple SCSI Target Ports are based
on
> multiple target portal groups as advertised by the target.
>
> On the other hand, there is no requirement of the target that if
advertise
> itself as having only one target portal group, even if it was capable.
It
> can subdivide its ipaddress space in any way it wants.   A high end
target
> with many many ip interfaces will probably do that.  Additionally, any
> truly high end target (in the long term) will have many iSCSI HBAs (most
> likely) each functioning as a target portal group and you'd see this
> modeling FC (at the target side) closer.
>
> There might be other reasons besides multiple iSCSI HBA configuration for
> the target to advertise multiple target portal groups. The SCSI
Asymmetric
> Port behavior for controllers in particular (for failover, primary
pathing,
> etc.) can take advantage of that kind of structure.  You may have a
> dual-headed controller each with independent power supplies.  They might
> have enough coordination to run sessions across all of them, but it might
> make more sense to separate them.  [Give me more time and I can probably
> come up with lots of other reasons too...]
>
> The model then is flexible enough to handle arbitrary software
> implementations using arbitrary network or TCP hardware cards as well as
> any implementations of iSCSI HBAs or any combination of the two. And
that's
> true on both the initiator and the target side.
>
> Jim Hafner
>
> Santosh Rao <santoshr@cup.hp.com>@ece.cmu.edu on 09/13/2001 05:37:45 pm
>
> Sent by:  owner-ips@ece.cmu.edu
>
> To:   IPS Reflector <ips@ece.cmu.edu>
> cc:
> Subject:  iscsi : target port definition
>
> Hello,
>
> I have a question on the interpretation of the iscsi target port
> definition. The iscsi rev 08 defines the iscsi target port to map to an
> iscsi target portal group.
>
> Thus, any iscsi target that wishes to allow multiple SCSI paths to be
> established to the target node MUST provide at least 2 iscsi target
> portal groups.
>
> The above definition of an iscsi target port somewhat alters the
> semantics of a target portal group. A target portal group, by
> definition, is a collection of a set of network portals within the
> target across which a session can be spanned.
>
> Thus, if a target supports a multi-connection session spanning across
> all its network portals, such a target would use a single target portal
> group to indicate that 1 big fat session pipe could be established to
> all its network portals. This, in turn, would have the side effect of
> only providing 1 scsi path to the upper layer wedge drivers, if the
> iscsi initiators establish a session per target portal group. [which is
> the target port].
>
> >From an initiator's perspective, what should be the target side
> end-point of an initiator's sessions when it may need to support upper
> layer wedge drivers ? Should the initiator establish a session per
> target
> portal group [, in which case the above issue exists] ? Or, should it
> establish a session per TargetAddress ??
>
> Regards,
> Santosh