Re: iSCSI: current UNH Plugfest

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

Attached are the new issues that arose during the iSCSI plugfest
at UNH on Wednesday 31-Oct-2001.

(Note: these issues do not take into account any modifications or
clarifications that occured in the standard due to the issues raised
on Monday or Tuesday.)

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

----------------------------------------------------------------------------

1. Are receivers (initiator or target) REQUIRED to check that reserved
   bits and/or fields are set to 0?

   Section 3 on page 48 of draft 8 says:
     "Any bits not defined MUST be set to zero.  Any reserved fields and
     values MUST be 0 unless specified otherwise."

   and Section 8.3 on page 127 of draft 8 says:
     "Explicit violations of the PDU layout rules stated in this document
     are format errors.  This when detected, usually indicates a major
     implementation flaw in one of the parties.

     When a target or an initiator receives an iSCSI PDU with a format
     error, it MUST reset all transport connections in the session
     immediately and escalate the format error to session recovery
     (section 8.11.4)."

   According to these rules, a PDU with reserved bits and/or fields that
   are not set to 0 violates the PDU layout rules.  Therefore, if an
   initiator or target receives such a PDU, it should immediately close
   all connections in the session and go to session recovery.

   Clearly a format error has extremely severe consequences!

   Although all vendors are setting reserved bits and fields to 0 on
   PDUs they are sending, many are NOT checking PDUs they are receiving
   to see if these bits and fields are set to 0.  Basically, vendors are
   saying "who cares if reserved bits and/or fields in incoming PDUs are
   not zero?  We do not want to take the time to do this checking, and
   there is no benefit to doing it.  As long as the non-reserved bits and
   fields are set properly, we can and should proceed.  Any time devoted
   to doing this checking is wasted in 99+% of the cases, and in the
   (unlikely) case that a non-zero bit or field is found, the
   consequences are too severe."

   There should be some statement in the standard to clarify what checking
   is required and what is optional.

2. A similar situation arises with respect to checking the consistency
   of fields such as Version-max, Version-min and Version-active in Login
   Requests and Login Responses.

   For example, consider the Version-max field.

   Section 3.12.5 says:
     "All Login requests within the Login phase MUST carry the same
     Version-max."

   All vendor initiators are setting Version-max correctly on all
   login requests they are sending, but many vendor targets are NOT
   checking received login requests to ensure that this rule is enforced.
   In particular, many targets simply use the Version-max and Version-min
   on the first login request they receive on a new connection, and then
   they ignore these fields on all subsequent login requests in the same
   login phase.

   Strictly speaking, a change in the Version-max field during the login
   phase constitutes a protocol error according to section 8.8 on page 130
   of draft 8:

     "All violations of iSCSI PDU exchange sequences specified in this
     draft are also protocol errors.  This category of errors can be
     addressed only by fixing the implementations; iSCSI defines Reject
     and response codes to enable this".

   Therefore the target should send back a login response with a status
   of 0x0200 and then close the connection.

   However, Section 3.12.5 also says:
     "The target MUST use the value presented with the first login request."

   This rule seems to imply that the value CAN change, because if it cannot
   change, then it doesn't matter which one of the login requests it is
   taken from, they are all the same anyway.

   The suggestion is to keep the requirement that the target MUST use the
   value presented on the first login request, but to allowed the target
   to ignore the value presented on all subsequent login requests in the
   same login phase.  A similar rewording should be done for the other
   fields.

3. Can commands be sent out of order on the same connection?

   The behavior of targets is clearly specified in Section 2.2.2.3 on
   page 25 of draft 8, which says:
     "Except for the commands marked for immediate delivery the iSCSI
     target layer MUST eliver the commands for execution in the order
     specified by CmdSN."

   Section 2.2.2.3 on page 26 of draft 8 also says:
     "- CmdSN - the current command Sequence Number advanced by 1 on
     each command shipped except for commands marked for immediate
     delivery."
   but the meaning of the term "shipped" is vague, and does not necessarily
   require that the PDUs arrive on the other end of a TCP connection
   in the same order that the CmdSN values were assigned to these PDUs.

   Some initiators have been designed to send commands out of CmdSN
   order on one connection.  Consider the situation where there is only
   one connection and a high-level dispatcher creates a PDU for a SCSI
   command that involves writing immediate data to the target.  This PDU
   is enqueued to a lower-level layer which has to setup, start, and
   wait-for a DMA operation to move the immediate data into an onboard
   buffer before the PDU can be put onto the wire.  While this is
   happening, the dispatcher creates another unrelated PDU for a SCSI
   read command (for example), and when this PDU is passed to the
   lower-level layer it can be sent immediately, ahead of the previous
   write PDU and therefore out of order on this connection.

   The standard clearly allows this to happen if the two PDUs were sent
   on different connections, and seems to imply that this can also happen
   when the two PDUs are sent on the same connection.

   The suggestion is to put in the standard an explicit statement that
   this is allowed or not allowed, as appropriate.
   
   If this is allowed, such a statement would avoid the erroneous
   assumption being made by some target implementers that within a single
   connection, commands will arrive in order.

   If this is not allowed, such a statement would avoid the erroneous
   assumption being made by some initiator implementers that within a
   single connection, commands can be put on the wire out of order.

4. Three numeric keys (MaxRecvPDULength, MaxBurstSize, FirstBurstSize)
   now allow: "A value of 0 indicates no limit."

   Is this useful?  Does it buy anything?

   The difficulties implementers are having with this are:

   1) It is a special case.
   2) It causes discontinuous ranges (for example, [0,64..2**24])
   3) It violates the min/max function normally used for the key.
   4) There is always a limit anyway.

   Consider FirstBurstSize, which can have a value that is described
   as "<0|number-64-2**24>", and for which the minimum of the 2 numbers
   is selected.

   I one side offers 0 to mean unlimited, and the other side
   has a limit, it will reply with that limit, say 128 Kbytes.
   Therefore, the result is not min(0, 128K) but rather max(0, 128K).
   The statement in the standard that "the minimum of the 2 numbers is
   selected" is therefore wrong when one of the numbers is 0.

   Furthermore, when an initiator or target receives an offer for one
   of these keys, it cannot simply check that the offered value is
   legal by testing it against some minimum and maximum.  It must first
   check for 0 and then only if that check shows the value is non-zero
   can it do the min/max range check for legality (i.e., 64-2**24).

   Finally, there is always a limit. If nothing else it is the
   limit imposed by the 24-bit DataSegmentLength field of the PDU
   requesting the transfer.  It is useless to specify a FirstBurstSize
   (or MaxRecvPDULength or MaxBurstSize) any bigger than that, because
   the largest possible DataSegmentLength in any PDU that can use
   this value is 2**24-1.

   The suggestion is to just eliminate this special case of 0 and require
   that the range 64-to-(2**24-1) be used instead -- it has exactly the
   same effect in all cases, it is easier to describe in the standard
   because it avoids all the extra words, and it is easier to code
   because it avoids all the special cases.

   NOTE: the standard should specify the limit in the ranges for
   MaxRecvPDULength, MaxBurstSize, and FirstBurstSize as 2**24-1 instead
   of 2**24.  The number 2**24 cannot be represented in the 24-bit
   DataSegmentLength field and therefore can never be used.

5. This is a suggestion for a minor rewording in the standard to avoid
   misunderstandings.

   In Appendix E on page 188 of draft 8 it says:

     "The response to this command is a text response containing a list of
     targets and their addresses.  Each target is returned as a target
     record.  A target record begins with the TargetName text key,
     followed by a list of TargetAddress text keys, ..."

   In fact, there are situations where there are no targets and/or no
   addresses.  These situations are clearly defined in the draft after
   the sentences quoted above, but it would help if those sentences
   at least hinted at the possibility that the lists could be empty
   or might not contain addresses.  A possible rewording would be:

     "The response to this command is a text response containing a list of
     zero or more targets and, optionally, their addresses.  Each target
     is returned as a target record.  A target record begins with the
     TargetName text key, followed by a list of zero or more
     TargetAddress text keys, ..."


6. This is a suggestion for another very minor rewording in the standard.

   At the end of section 2.2.3 on page 29 of draft 8 it says:

     "Before full feature phase is established, only Login PDUs are
     allowed. ..."

   The suggested rewording is:

     "Before full feature phase is established, only Login Request and
     Login Response PDUs are allowed. ..."