Re: Bidirectional SCSI commands and iSCSI

[julian_satran@il.ibm.com]

Jim,

I'll take a look at the proposals (I assume they are on the T10 site).
My concern about cluttering is related to the attempt to make a
fixed-length header
as I think that software initiators will be widely used and they are far
more efficient with fixed length headers for widely used operations.

Thanks,
Julo

"Jim Hafner/Almaden/IBM" <hafner@almaden.ibm.com> on 02/10/2000 18:22:40

Please respond to "Jim Hafner/Almaden/IBM" <hafner@almaden.ibm.com>

To:   Julian Satran/Haifa/IBM@IBMIL
cc:
Subject:  Re: Bidirectional SCSI commands and iSCSI





Julo,

True, those examples are not true duplex operations. These are only the
current thinking or motivation for bi-directional commands.  They are not
the end of the story, I'm sure.

However, the only SAM-2 requirement on bi-directional commands (as now
defined) is that DataOut buffer be ready at the time the command is issued
(so the target can get/request that data at any time after receipt of the
command).  In other words, the content of the DataOut buffer should not be
time or data-dependent on the contents of the DataIn buffer.  This means
that as far as SAM-2 is concerned, there is no problem with a true duplex
operation.  (We just don't have any examples at the moment.)

I think your second proposal (using "immediate" data) is only attempting to
address the specific examples (OSDs), and not the generic notion of
bi-directional commands.  How would that solution work if the DataOut
buffer was a 1MG chunk of user data for a XDREADWRITE command?

The first method (splitting, if I read it correctly) undercuts the
motivation that is generating this bi-d command in the first place. The
SCSI layer could do that as well, but the issue is one of atomicity of
command.  (Aside: one principle governing definitions of bi-directional
commands in T10 is that there must always be a two (or more)
uni-directional variant of the command, to allow a complex operation to be
performed on a traditional uni-directional with simpler steps. I read your
suggestion as having iSCSI do this for the SCSI layer -- there'd be no
point.  Besides, doesn't this break layering?).

Finally, I would suggest a look at the FCP-2/3 proposal just submitted. I
think they are proposing (for legacy reasons) a new IU for bi-directional
commands.  If you're concerned about cluttering your header, then this
group might think about adding a new PDU for these types of commands.

I don't quite understand the reluctance to add new fields in the header for
this purpose.  The current design is not yet cast in stone and now is the
perfect opportunity to get it defined in a way that accomodates the future
that T10 seems to be moving (rapidly) towards.

Jim Hafner


julian_satran@il.ibm.com@ece.cmu.edu on 10-01-2000 12:16:06 PM

Sent by:  owner-ips@ece.cmu.edu


To:   ips@ece.cmu.edu
cc:
Subject:  Re: Bidirectional SCSI commands and iSCSI





Jim,

Both the examples you gave have something peculiar - the operations are to
executed really in sequence (we are no talking about full duplex I/O
operations?).
I any case, in order no to extend excessively the headers we can either
introduce
a "next header" field (as in IP V6) or split the operation at the iSCSI
level into
two phases - each carrying its parameters and each ended with a status -

like:

  - out XOP1, XOP2, OD1,
  -in S1, ID, S2



In any case commands having only a parameter set (fixed or variable) - like
those
for the object disk should be handled with the parameters sent as immediate
data
(as with the three party commands) - parameter size will probably be
limited
and the performance requirement will be similar to those of block
read/write.
Do we need for those to care also about residual counts for parameters?


Julo

"Jim Hafner/Almaden/IBM" <hafner@almaden.ibm.com> on 29/09/2000 00:21:32

Please respond to "Jim Hafner/Almaden/IBM" <hafner@almaden.ibm.com>

To:   Julian Satran/Haifa/IBM@IBMIL
cc:   ips@ece.cmu.edu
Subject:  Re: Bidirectional SCSI commands and iSCSI





Julo,

The primary motivation is for object-based storage devices (OSDs).  The
protocol that seems to be needed in that environment is the ability to send
a  CDB, plus some parameter data (DataOut) AND get back read-type data
(DataIn) in the same command sequence.

One (non-OSD) command has been proposed in T10 for an XOR type XDREADWRITE
command.

SAM-2 has now allowed for such commands in the architecture. From this
perspective, it is completely symmetric. Transports that support this
feature will need to have data structures in there PDUs or IUs (or
whatever) to indicate things like DataOut datalength, DataIn datalength,
possible virtual addresses for each buffer, etc.  In other words, there
will potentially be two active buffers (on each end) for use with the same
command.    So, e.g., iSCSI PDU may need to have both "DataIn expected data
transfer length" and  "DataOut expected transfer length" fields.  I'm
assuming that the data direction in the existing iSCSI PDU is derived from
the SCSI opcode, unlike FCP where there are fields to indicate the data
direction.  (I am surprised not to find such things in the iSCSI Command
PDU.)

There are two proposals in T10 to enable this feature in SPI-4 and in FCP-x
(probably FCP-3).

Jim Hafner


julian_satran@il.ibm.com@ece.cmu.edu on 09-28-2000 12:10:53 PM

Sent by:  owner-ips@ece.cmu.edu


To:   ips@ece.cmu.edu
cc:
Subject:  Re: Bidirectional SCSI commands and iSCSI





Jim,

What is the intended use?  Are there already commands that use it?
Is there a major/minor distinction - like output data & input results or
output
parameters & input data - or is it intended to be completely symmetric?

Julo

"Jim Hafner/Almaden/IBM" <hafner@almaden.ibm.com> on 20/09/2000 17:56:27

Please respond to "Jim Hafner/Almaden/IBM" <hafner@almaden.ibm.com>

To:   ips@ece.cmu.edu
cc:    (bcc: Julian Satran/Haifa/IBM)
Subject:  Bidirectional SCSI commands and iSCSI




Folks,

In case you missed this, T10 just approved changes to SAM-2 (rev14) which
remove the restriction that a single SCSI command can have a data flow in
only one direction (DataIn xor DataOut).   Transport specs (that want to
allow this) will need to change to allow for two sets of Data buffer fields
in the headers (length, over/underrun, etc.).

Is anybody looking to modify the iSCSI spec to get these concepts
supported?

Jim Hafner