Re: iSCSI: CmdSN during the Login phase

["Robert D. Russell" <rdr@mars.iol.unh.edu>]

Mark:

I respectfully disagree with your recommendation to go with option 3.

The experience here at the plugfest seems to be that option 1 is
easiest to work with since it involves no special case code --
the cmdsn is incremented on every command sent.  Period.
A simple statement to that effect in the standard is much, much
prefered than a long chain of "technical" reasoning that readers have to
follow through to the end to infer what the final result should be.

I believe simplicity is the key to interoperability, and the standard
mechanisms should be designed to be simple.  There is indeed a reason
to number commands during login -- because then they are treated the
same as any other commands and you don't need special purpose code
during login to deal with that part of shipping the commands.
I.e., you keep it uniform and simple, no special cases.
This allows code to be reused and debugged once.

As a final point, most implementations will NOT have to be modified
to implement option 1 -- they now already do that.

Bob Russell
InterOperability Lab
University of New Hampshire
rdr@iol.unh.edu
603-862-3774

On Tue, 17 Jul 2001, Mark Bakke wrote:

> 
> Given that this is being interpreted differently by enough
> implementations, I would prefer to see us take a step back
> and look at which of Scott's three options should make the
> most sense.  It looks like options #1 and #2 are what is
> currently being implemented, regardless of what's specified.
> 
> However, a connection technically can't really be part of a
> session (sharing its work load) until it has completed a
> successful login phase.  There's also little reason for
> numbering until the login phase is complete, since the login
> and text responses during negotiation are synchronous, and
> pertain only to the connection, rather than the session.
> 
> I would recommend stating that a connection joins (or becomes,
> if it's the first one) a session after the login response
> with the F bit set is sent (on a target), or after the login
> response with the F bit set is received (on an initiator),
> and that CmdSN is not necessary until such a time, especially
> since CmdSN is per-session, not per-connection.  This is 
> identical to Scott's option 3.
> 
> I think that defining it this way would remove any further
> ambiguity on when a connection is part of a session.  Since
> it appears that most implementations will have to modify their
> login code anyway, this shouldn't be too bad.
> 
> --
> Mark
> 
> Matt Wakeley wrote:
> > 
> > I think it's pretty obvious that it's your option #1: (the initiator supplies
> > the initial CmdSN in the login PDU, and for every command PDU after that, the
> > CmdSN gets incremented, just like in "full feature phase")
> > 
> > 2.10.7 CmdSN says "CmdSN is either the initial number of a session..."
> > [Julian, this should say "initial command number of a session"]
> > 
> > So, the Login PDU defines the first CmdSN.
> > 
> > 1.2.2.1 says "Commands in transit from the initiator to the target layer are
> > numbered by iSCSI; the number is carried by the iSCSI PDU as CmdSN".  Text
> > commands are commands, so they are numbered also.
> > 
> > It goes on to say "CmdSN - the current command Sequence Number advanced by 1
> > on each command shipped except for commands marked for immediate delivery."
> > 
> > I agree that this needs to be cleaned up. CmdSN and StatSN should work during
> > login just like they do during "full feature phase" - there is no reason why
> > the should not, and making them always work the same removes complexity and
> > interoperability problems.
> > 
> > -Matt
> > 
> > "Scott M. Ferris" wrote:
> > >
> > >   At what point during the Login Phase of a leading connection (null
> > > TSID) does a session exist?
> > >
> > >   Section 2.10.7 describing CmdSN for the Login PDU indicates that
> > > when TSID is null, CmdSN indicates the starting Command Sequence
> > > number for this session.
> > >
> > >   For interoperability, it is important to define precisely when a
> > > session exists, so that all initiators and targets agree on when
> > > CmdSNs are significant, and do not reject or ignore PDUs due to
> > > differing assumptions of how the CmdSN numbering should be done during
> > > the Login Phase.
> > >
> > > Some of the possible choices for session start are:
> > >
> > > 1) a session starts before the initiator sends anything, and the Login
> > >    PDU is the first PDU of a session.
> > >
> > > 2) a session starts when the initiator receives a (possibly partial)
> > >    Login Response with an accept status class and a non-zero TSID, and
> > >    the next PDU sent by the initiator is the first PDU of the session,
> > >    which uses the same CmdSN as the Login PDU.
> > >
> > > 3) a session starts when the initiator receives a Final Login
> > >    Response, and the next PDU sent by the initiator is the first PDU
> > >    of the session, which uses the same CmdSN as the Login PDU.
> > >
> > >   After searching through draft 6, I was unable to find anything that
> > > clearly indicated which if any of the above possibilities was correct,
> > > or, if more than one is possible, how the initiator and target are
> > > supposed to determine what CmdSN numbering scheme the other side is
> > > using for the Login Phase.
> > >
> > >   The UNH draft-6 initiator appears to use choice #1.  The Cisco
> > > draft-6 initiator and target use choice #2.  This can cause the UNH
> > > initiator to hang when trying to login to the Cisco target, since the
> > > initiator may send a Text PDU with CmdSN 2, while the target is
> > > waiting for a PDU with CmdSN 1.
> > >
> > >   Section 1.2.2.2 states that "Status numbering starts after
> > > Login. During login, there is always only one outstanding command per
> > > connection and status numbering is not strictly needed but may be used
> > > as a sanity check."
> > >
> > >   A similar argument could be made that CmdSN is not needed during the
> > > Login Phase, suggesting choice #3 may be the simplest.
> > >
> > >   I also think section 1.2.2.2 is problematic, as it appears to
> > > contradict itself by saying that status numbering is not used during
> > > login, and then follows up saying status numbering may be used during
> > > login.
> > >
> > > --
> > > Scott M. Ferris,
> > > sferris@acm.org
> 
> -- 
> Mark A. Bakke
> Cisco Systems
> mbakke@cisco.com
> 763.398.1054
>