Re: iSCSI: MaxRecvPDULength simplification

["Mallikarjun C." <cbm@rose.hp.com>]

One of the design goals of task reassign is to be able to reassign the
connection allegiance of a task to a different connection that's perhaps
using an entire different network for a true failover capability - for ex.,
the original allegiant connection could be traversing a corprorate intranet,
and the reassigned connection could be using the public internet.  That's
the reason for not placing the equivalence requirement on the two
connections.

MaxRecvPDULength was made negotiable during the FFP because
several of us felt that it's very useful for iSCSI to allow implementations
with the `ULPDU containment' property
(http://search.ietf.org/internet-drafts/draft-ietf-tsvwg-tcp-ulp-frame-01.txt).
This automatically means that MaxRecvPDULength should be changeable
with changes in PMTU changes, particularly with dynamic PMTU degradation.

Let's look at the possibility you suggest - there being two valid MaxRecvPDULength
values at the same time.  If a change in FFP is being attempted for this key, it must
be due to PMTU changes, and the implicit expectation is that initiator will sense
it and initiate the negotiation process (thus enabling the target to declare its changed
MaxRecvPDULength, if any, as well in the text response).  There is no ambiguity for
the initiator since the text negotiation can not be assumed successful (meaning new
MaxRecvPDULength can't be used) until the text response (with F-bit) is received.
As far as the target is concerned, it can not assume the text negotiation to be effective
(meaing new MaxRecvPDULength can't be used) until the StatSN for the text response
is ack'ed - perhaps we should make this clear in the text (and allow explicit soliciting
for StatSN acknowledgement using a NOP-ping).
--
Mallikarjun

Mallikarjun Chadalapaka
Networked Storage Architecture
Network Storage Solutions Organization
Hewlett-Packard MS 5668
Roseville CA 95747
cbm@rose.hp.com

----- Original Message -----
From: "Michael Schoberg" <michael_schoberg@cnt.com>
To: "IPS Reflector (E-mail)" <ips@ece.cmu.edu>
Sent: Wednesday, March 27, 2002 8:36 AM
Subject: RE: iSCSI: MaxRecvPDULength simplification


> SNACK with RL=0 is only one requirement.  Another is task allegiance
> reassignment (6.1.2) which does not use SNACK (correct?)  Here the initiator
> sends a Task-Management request with a Task-Reassign message to the target.
> The target must reorganize it's T->I messages to match the MaxRecvPDULength
> of the new connection.  Plus, the Task Reassign message includes the
> ExpDataSN field which, if I'm reading right, on reassignment the target must
> remember the sequencing of the old connection in order to track the
> ExpDataSN field then re-sequence for the new connection (9.5.4).  So now
> target implementations will have to keep track of sequence indexes for
> Data-In PDUs for conversion over to the new allegiance.
>
> On-the-fly modification of MaxRecvPDULength also means that some compliance
> standard must be set for in-flight messages.  An outstanding read request
> may have T->I PDUs that comply with multiple MaxRecvPDULength values.  In a
> sense, there will be times where multiple MaxRecvPDULength values are valid.
> At what point MUST the target comply with the new value?
>
> So I guess I'm not convinced that the benefit of simplification is
> illusionary.  Is there a discussion thread that expressly states
> MaxRecvPDULength must have the flexibility allowed for in the draft?  The
> discussions I've seen mainly centered around whether it should be duplex
> valued (I->T, T->I).
>
> Thanks.
>
> -----Original Message-----
> From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
> Sent: Wednesday, March 27, 2002 1:40 AM
> To: Michael Schoberg
> Cc: IPS Reflector (E-mail); owner-ips@ece.cmu.edu
> Subject: Re: iSCSI: MaxRecvPDULength simplification
>
>
>
> The simplification you are talking about is illusory. Currently this
> requires SNACK to require all data(RL = 0).
> The restriction you propose instead is far more severe (and unwarranted).
>
> Julo
>
>
>
> Michael Schoberg <michael_schoberg@cnt.com>
> Sent by: owner-ips@ece.cmu.edu
>
>
> 27-03-02 01:19
> Please respond to Michael Schoberg
>
>
>
>         To:        "IPS Reflector (E-mail)" <ips@ece.cmu.edu>
>         cc:
>         Subject:        iSCSI: MaxRecvPDULength simplification
>
>
>
>
> I've looked over some of the reflector messages regarding MaxRecvPDULength
> negotiation (back to Oct 2001).  I can't find the discussion of why this
> value must be available for negotiation outside of login.  It really
> complicates SNACK and Task Reassignment if the initiator can change this
> value on-the-fly.  Would it be too restrictive to propose the following
> rules?
>
> 1) MaxRecvPDULength is negotiated only at login before FFP.
>
> 2) For message-level recovery, a connection must be prepared to accept the
> MaxRecvPDULength of the connection it's providing fail over capability for.
> It doesn't seem unreasonable to expect fault tolerant configurations to
> comply with this.  It would simply be stating that RecoveryLevel 2 devices
> should be somewhat matched in this capability.
>
>
>
>  -----Original Message-----
> From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
> Sent: Tuesday, March 26, 2002 1:50 AM
> To: Michael Schoberg
> Cc: IPS Reflector (E-mail); owner-ips@ece.cmu.edu
> Subject: Re: iSCSI: SNACK RunLength
>
>
> Michael,
>
> The paragraph says that if you issue a SNACK after the MaxPDURecvLength has
> decreased you should use RunLength 0 (meaning all) instead of a specific
> runlength which might not be accurate anymore.
>
> Julo
>
>
> Michael Schoberg <michael_schoberg@cnt.com>
> Sent by: owner-ips@ece.cmu.edu
>
>
> 25-03-02 19:43
> Please respond to Michael Schoberg
>
>
>        To:        "IPS Reflector (E-mail)" <ips@ece.cmu.edu>
>        cc:
>        Subject:        iSCSI: SNACK RunLength
>
>
>
>
>
> Am I just not reading this or can this paragraph use some help?  Can someone
> give an interpretation?
>
>  9.16.3  RunLength
>                                 ...
>
>          The first data SNACK, issued after initiator's MaxRecvPDULength
>          decreased, for a command issued on the same connection before the
>
>          change in MaxRecvPDULength, MUST use RunLength "0" to request
>          retransmission of any number of PDUs (including one).  The number
> of
>          retransmitted PDUs in this case, may or may not be the same as
> the
>          original transmission, depending on whether loss was before or
> after
>          the MaxRecvPDULength was changed at the target.
>
>
> Does this refer to something like:
>
> For connections with unacknowledged PDUs that undergo MaxRecvPDULength
> negotiation ...
>
>
>
>
>
>
>