iSCSI: SessionTypes

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

Folks,

As was noted in another thread, the recent drafts (07 included) introduced
a new mechanism for avoiding the issue of login to the "default target" for
SendTargets only function.   It did this with the "SessionType" key.  I
like this idea a lot.

However, the draft proposes two additional session types besides
"Discovery" (for SendTargets only) and "Normal" (for SCSI to a real live
target).  The two additional ones are "Boot" and "CopyManager" and in both
of these additional cases, it is suggested that the target might limit what
SCSI commands are allowed.

I have very strong feelings that these are (a) both unnecessary, (b)
strongly violate layering AND (c) are incompletely specified, in any case.

It's unnecessary because in both cases, the intent seems to be to limit
what SCSI commands might be allowed within the given session, but if an
initiator voluntarily requests such limiting behavior, then it can
voluntarily limit what SCSI commands it sends.  For the initiator to ask
for a filter from the target when it can filter itself is silly.

With respect to layering, this would be the first protocol that *might*
restricts the set of SCSI commands allowed. In affect, it allows the iSCSI
layer to filter the SCSI layer by changing the set of commands supported by
a particular device type.  That could get very confusing for the SCSI layer
in the initiator (it sends a command and the iSCSI target layer rejects it,
even though the device should support the command).  It is also well beyond
what a protocol spec should do.

The proposal does not say what error conditions are reported if a command
is rejected.  By saying it's "vendor-dependent", it leaves the door open
for massive interoperability problems (one target doesn't filter, another
filters most everything).

I can possibly foresee iSCSI specific reasons for such things (e.g., to
request different authentication methods or security context, in analogy
with Discovery session type), but until those are defined in detail, I see
no reason to keep these things in.  At best they might be reintroduced in
the second generation of the standard.

Consequently, I would *strongly* suggest that these two be removed from the
draft.

Comments?
Jim Hafner