Re: iscsi : changes involving tgt portal group tag.

[Santosh Rao <santoshr@cup.hp.com>]

Jim & Mallikarjun,

I believe the following 3 rules should suffice to prevent any terminated
& re-established sessions from delivering I/Os to incorrect target ports
and/or LUNs :

1) The TPGT is sent as a field in the login reqest PDU. Target MUST
reject login if the wrong TPGT (target port) is being addressed. 

2) TPGTs MUST NOT be re-used when new target ports are instantiated. If
the entire 64K space has been used up for TPGTs, then, the target node
name must be changed. (May be worth considering using a 32-bit field or
a 64-bit field for the target portal group).
This rule is needed, since port names are required to be world-wide
unique, and re-use of TPGTs upon repeated target re-config will cause
them to lose world-wide unique-ness. 


3) Any time a portal group is re-configured such that the LU view behind
the portal group (target port) has changed, the target MUST generate a
"REPORTED LUNS DATA HAS CHANGED" check condition. 

Jim, in response to your comments :

I agree that the change in LUN inventory behind a target port is best
detected and reacted to on a "REPORTED LUNS DATA HAS CHANGED" check
condition.

As for using the LUN WWNs (EVPD INQUIRY to get page 83h) in order to
authenticate all the LUNs behind a target port, I think the above rules
are a more efficient way to detect this on session termination &
re-establishment, rather than stall all I/O traffic to that target port,
while going out to each LUN to authenticate them based on their LUN
WWNs. Besides, you still have to deal with the case where the LUN does
not support LUN WWNs.

Regards,
Santosh



Jim Hafner wrote:
> 
> * From the T10 Reflector (t10@t10.org), posted by:
> * "Jim Hafner" <hafner@almaden.ibm.com>
> *
> 
> Mallikarjun,
> 
> Much as I would like to agree with you on most things (especially after you
> agreed with me so much...)
> 
> I think connecting LU views with TPGTs is a mistake and violates layering.
> Besides, LU views are
> (a) functions of SCSI level or other access controls
> (b) functions of iSCSI target node presentations
> Within the standards that I know of, there is no defined LU view that is
> different across the different target ports of any SCSI device (BUT see the
> two notes in the P.S.)
> 
> In case (b), the TPGT is already qualified by the node so there's no issue
> here.
> In case (a), the rules as defined by SCSI ACs are that the LU view is the
> same across all target ports.
> 
> I don't have any good "words" to smith into the document that makes clear
> what "reconfiguration" should trigger a new name and what shouldn't (and I
> haven't had time to think much about it either).
> 
> As for your final comment about IOs getting directed to the correct LU.
> This should be handled by LU Identifier (not just LUN) if this is a
> significant concern.  There's a subtle requirement in this virtual world
> we're moving towards that requires careful mapping of LUs to device objects
> within hosts so that LUN changes don't cause such problems.  But LUN
> changes can occur for many reasons, not just those raised here.  SCSI gives
> "REPORT LUNS DATA CHANGED" Unit Attentions, but if they're ignored,
> ...OOPS!.
> 
> Finally, to Santosh:  there is no notion in SCSI (or iSCSI, I think) about
> LUN OPEN/CLOSE.   LUs are "discovered" by INQUIRY and REPORT LUNs.  Once
> discovered, there is no further action required.  So, a logical unit is
> either "addressible" (in the view) or not (with a slight caveat about SCSI
> ACs and AccessId registration, but that should happen before the discovery
> phase anyway).   There may be layers about SCSI which have these notions
> but they probably only manipulate internal OS datastructures and have no
> "wire" protocol associated with them.   So, there should be no need to
> think about such actions being required after re-login.
> 
> Jim Hafner
> 
> P.S.
> Note1:  there are vendor-specific implementations that might seem to
> violate this statement if taken literally, but because such implementations
> are "pre-node names", one can view them as falling into the model (b)
> above.
> Note2: the "asymmetric port" stuff in SCSI (recent addition) doesn't change
> the "view" (i.e., the list of LUNs) so much as the accessibility of the
> addressed LU, so even this doesn't square with the issue at hand.
> 
> "Mallikarjun C." <cbm@rose.hp.com>@t10.org on 03/04/2002 03:12:15 PM
> 
> Sent by:    owner-t10@t10.org
> 
> To:    "Santosh Rao" <santoshr@cup.hp.com>, <ips@ece.cmu.edu>
> cc:    <t10@t10.org>
> Subject:    Re: iscsi : changes involving tgt portal group tag.
> 
> * From the T10 Reflector (t10@t10.org), posted by:
> * "Mallikarjun C." <cbm@rose.hp.com>
> *
> Santosh and Jim,
> 
> It appears a good idea to add in the portal group tag in the Login
> Request header for a sanity check by the receiving target portal.
> 
> I generally agree with Jim's comments that the duration of persistence
> of a portal group tag is intricately linked to the extent of portal
> reconfiguration.
> Not every target reconfiguration may result in a portal group tag
> reassignment.
> OTOH, some reconfigurations may be so extensive as to cause even a node
> name change.
> 
> Some comments on Santosh's message -
> 
> > "If a portal group is re-configured such that all its previously
> > advertised network portals are no longer a part of the portal group,
> > then, the portal group tag (and thereby, the port name/identifier)
> > *MUST* be changed to indicate a new target port."
> 
> I am not sure this solves the problem you're trying to get at.  If none of
> the earlier IP addresses can get an initiator to the SCSI target port that
> it
> knew of (your scenario), it appears to me that it doesn't matter if the
> portal group tags are changed or not....A new discovery process should
> update the initiator of the changed portal addresses.
> 
> I suggest the following text -
> 
>    After a portal group reconfiguration which changed the view of LUs
>    for an initiator with a given set of privileges, the target MUST change
>    the portal group tag that represents the reconfigured target portal
>    group.
> 
> > > Under these events, shouldn't all "open/active I_T_L traffic" have been
> > > aborted, closed or otherwise ended?  So this shouldn't be an issue at
> all.
> >
> > On a session logout & re-login, it is not efficient/necessary to close
> > and re-open each LUN behind that target port, due to the fact that a
> > target port may have hundred's of LUNs behind it.
> 
> I agree with Jim on this one - there should be *no* open/active I_T_L nexus
> traffic after a reconfiguration, targets can enforce this via simple iSCSI
> means
> (reject initiator advances to continue the session for DefaultTime2Wait+
> DefaultTime2Retain seconds).  In fact, Async logout request requires a
> clean closure that implicitly aborts I/Os.
> 
> What you're describing is typical O/S "LUN open" and "LUN close"
> interactions.  I agree that they wouldn't have to be repeated, but care
> must be taken to ensure that new I/Os (on the new session after
> reconfiguration)
> are not delivered to the incorrect LUs.  It seems that the addition of
> TPGT in the login header and the proposed new text (above) would take
> care of this.
> 
> Regards.
> --
> Mallikarjun
> 
> Mallikarjun Chadalapaka
> Networked Storage Architecture
> Network Storage Solutions Organization
> Hewlett-Packard MS 5668
> Roseville CA 95747
> 
> ----- Original Message -----
> From: "Santosh Rao" <santoshr@cup.hp.com>
> To: <ips@ece.cmu.edu>
> Cc: <t10@t10.org>
> Sent: Monday, March 04, 2002 10:40 AM
> Subject: Re: iscsi : changes involving tgt portal group tag.
> 
> > * From the T10 Reflector (t10@t10.org), posted by:
> > * Santosh Rao <santoshr@cup.hp.com>
> > *
> > Jim,
> >
> > I agree that a "complete re-configuration" of a target port can result
> > in a new port name & port identifier. However, the tricky part is the
> > definition of a "complete re-configuration of an iscsi target port", due
> > to the concepts of portal groups involving multiple network portals.
> >
> > For example, if a portal group (aka, an iscsi target port) were to be
> > re-configured to include a new network portal [moved from another portal
> > group], then, the target port itself has not changed, since it is still
> > accessible through its previously used network portals. What has changed
> > is the individual network portal that has moved from 1 target port to
> > another.
> >
> > Hence, it is sufficient, in this case, to maintain persistence of the
> > target port name/identifier, without requiring any change in port
> > name/identifier. By requiring initiators to send the intended TPGT (scsi
> > target port name/identifier) along with the login request, this allows
> > the target port to detect that the network portal is being accessed to
> > connect to a different target port and it can reject the login.
> >
> > It may be helpful to call out the specific case when a port
> > name/identifier MUST change. How about something like :
> >
> > "If a portal group is re-configured such that all its previously
> > advertised network portals are no longer a part of the portal group,
> > then, the portal group tag (and thereby, the port name/identifier)
> > *MUST* be changed to indicate a new target port."
> >
> > This would allow access to the target port through its un-altered
> > network portals to continue un-disrupted. More comments inline, in
> > response to some of your queries.
> >
> > Thanks,
> > Santosh
> >
> > NOTE : In this discussion target port and target portal group are used
> > to imply the same entity, within a target node.
> >
> >
> > Jim Hafner wrote:
> >
> > > SAM-2 requires scsi port names to be persistent and world-wide-unique.
> > > (SAM-2 Rev 22 Section 4.7.7). Quoting from SAM-2 :
> > >
> > > "A scsi port name shall never change and may be used to persistently
> > > identify a scsi initiator port or target port...".
> > >
> > > <JLH>
> > > There are different ways that one can interpret this "persistent" rule.
> The
> > > intent was that names shouldn't change over time when *identifying the
> same
> > > object*.   What that means is that if the object changes (in any
> > > discernable way), then the name should change.  So, the object can move
> to
> > > a different piece of hardware, but it can still be named the same way.
> If
> > > the object changes, e.g., in this case, reconfigures as a different set
> of
> > > network portals with different addressing elements, then I would think
> that
> > > the name should change as well.   That's all the persistence one really
> > > needs.
> > >
> > > I'm not sure what that implies about your recommendation.  I don't see
> any
> > > problem with it, either.
> > > </JLH>
> >
> > I think we may be in agreement. (See reasoning above).
> >
> >
> > >
> > > The rationale for (2) is :
> > > --------------------------
> > > Consider an initiator node establishing multiple sessions to a scsi tgt
> > > port, with each session established to a subset of the network portals
> > > within the tgt portal group.
> > >
> > > Consider an iscsi transport event involving the re-configuration of
> > > target portal groups on the iscsi target node. This may be preceeded by
> > > the iscsi sessions seeing an async message "target requests logout",
> > > followed by session logout, portal group re-configuration, and then,
> the
> > > initiator re-establishes sessions.
> > >
> > > Across a transport event that results in such session termination and
> > > re-establishment, the initiator needs to authenticate that it is still
> > > speaking to the same [i]scsi target port, in order to ensure that any
> > > open/active I-T-L nexus traffic on that session is not incorrectly
> > > routed to a wrong LUN after such a transport event.
> >
> > > <JLH>
> > > Under these events, shouldn't all "open/active I_T_L traffic" have been
> > > aborted, closed or otherwise ended?  So this shouldn't be an issue at
> all.
> >
> > On a session logout & re-login, it is not efficient/necessary to close
> > and re-open each LUN behind that target port, due to the fact that a
> > target port may have hundred's of LUNs behind it.
> >
> > All outstanding I/Os need to be aborted. Once the session is
> > re-established [and the target port is authenticated to be the same],
> > further I/O traffic can be resumed without requiring the SCSI ULP to
> > close and re-open each LUN. Some transport specific clearing effects may
> > have occurred, which may require additional LUN level activity, in some
> > cases.
> >
> > (There are analogies to the above model in FCP & SRP, which authenticate
> > port name/identifier on login, allowing I/O resumption after session
> > termination and re-establishment.)
> >
> >
> > > To prevent such authentication issues, iscsi can send the iscsi target
> > > port identifier (portal group tag) explicitly in the login request, and
> > > the login can be rejected by the target on a portal group tag
> mis-match.
> > > (if the network portal does not belong to the addressed portal group
> > > tag).
> > > <JLH>
> > > Did you mean for the initiator to send this TPGT?
> > > </JLH>
> >
> > Yes. The initiator login request must include the target portal group
> > tag, thus identifying the target port to which a session is being
> > established.
> >
> > Login currently carries no addressing information, since the addressing
> > is implicit, based on the TCP connection in use. The problem with this
> > approach is that the login implicitly addresses a network portal, and
> > not the target port. As seen in the earlier example, the association of
> > network portal <-> target port can change, and such changes can be
> > detected, if the initiator were to explicitly identify the target port
> > being logged into.
> >
> > --
> > ##################################
> > Santosh Rao
> > Software Design Engineer,
> > HP-UX iSCSI Driver Team,
> > Hewlett Packard, Cupertino.
> > email : santoshr@cup.hp.com
> > Phone : 408-447-3751
> > ##################################
> > *
> > * For T10 Reflector information, send a message with
> > * 'info t10' (no quotes) in the message body to majordomo@t10.org
> >
> *
> * For T10 Reflector information, send a message with
> * 'info t10' (no quotes) in the message body to majordomo@t10.org
> 
> *
> * For T10 Reflector information, send a message with
> * 'info t10' (no quotes) in the message body to majordomo@t10.org

-- 
##################################
Santosh Rao
Software Design Engineer,
HP-UX iSCSI Driver Team,
Hewlett Packard, Cupertino.
email : santoshr@cup.hp.com
Phone : 408-447-3751
##################################