RE: iSCSI: MaxCmdSN

["Ayman Ghanem" <aghanem@cisco.com>]

Nick,

Yes, it will be an error for MaxCmdSN to have a value < (ExpCmdSN - 1), and
I agree that MaxCmdSN = (ExpCmdSN - 1) should be permitted. As for the
target receiving a PDU with CmdSN > MaxCmdSN, it is stated in draft-06 (near
the end of 1.2.2.1) that the target MUST silently ignore such a command. No
reject PDU needs to be sent.

-Ayman


> -----Original Message-----
> From: Martin, Nick [mailto:Nick.Martin@compaq.com]
> Sent: Monday, June 11, 2001 12:00 PM
> To: 'Ayman Ghanem'; ips@ece.cmu.edu
> Subject: RE: iSCSI: MaxCmdSN
>
>
> Ayman,
>
> I thought I understood this, and I had a great explanation supporting the
> draft (06), but I now realize I was "off by one".  I believe you
> are correct
> that ExpCmdSN will be one greater than MaxCmdSN if-and-only-if an
> acknowledgement is to be sent while the window is closed.
>
> The purpose of ExpCmdSN and MaxCmdSN transmitted by the target, is to
> advance values with the same names maintained in the initiator.  ExpCmdSN
> acknowledges receipt of all CmdSN values < ExpCmdSN.  MaxCmdSN is the
> largest CmdSN the target promises to accept, and conversely the largest
> CmdSN the initiator is permitted to send.
>
> The target can not regress these numbers in the initiator.  When
> the target
> transmits (and the initiator accepts) N as the MaxCmdSN, the initiator is
> allowed to transmit CmdSN values through N.  Then it must stop.
> Any number
> of additional PDUs may carry the same MaxCmdSN value.  None of these will
> cause the initiator's value to advance.  A response carrying a
> MaxCmdSN less
> than the initiator's current MaxCmdSN value is presumed to have arrived
> late.  It does not advance nor regress the initiator's present value.
>
> After the initiator sends CmdSN of MaxCmdSN (still N), it can not send any
> more until MaxCmdSN advances.  It should be permitted for the target to
> acknowledge the receipt of all commands received through this
> CmdSN which is
> N.  In this case ExpCmdSN would be N + 1 and MaxCmdSN is still N.  We know
> that no commands are in flight.  When the target is able to accept another
> command, it must advance MaxCmdSN (value > N in serial arithmetic), and it
> will send the new value to the initiator in the next inbound PDU (i.e.
> data-in, SCSI response, or nop-in).
>
> Yes, I believe it would be a protocol error for MaxCmdSN to have a value
> less than ExpCmdSN - 1.  This would be an acknowledgement for a
> CmdSN which
> the initiator was not yet permitted to send.  I think MaxCmdSN equal to
> ExpCmdSN - 1 should be permitted.
>
> Note however that nothing will break if this acknowledgement is not
> immediately sent or, having been sent, does not advance the initiator's
> values.  The window is already closed and remains so.  While the window is
> closed, the initiator will be holding the last PDU sent as unacknowledged,
> when the target has in fact received it.  The PDU will be successfully
> acknowledged when MaxCmdSN advances, because ExpCmdSN in the
> target will not
> advance simultaneously while the window is closed.
>
> Was this done intentionally to delay the last acknowledgement when the
> window is closed, or is it an "off by one" issue in  draft 06?
>
> This made me think about a related issue:
> It think it would be permitted for the target to reject or drop a PDU with
> CmdSN greater than MaxCmdSN, but I am not sure if it is required to do so.
> I do not see a response code the target could send that would tell the
> initiator what it did wrong.
>
> Thanks,
> Nick
>
> -----Original Message-----
> From: Ayman Ghanem [mailto:aghanem@cisco.com]
> Sent: Monday, June 11, 2001 9:18 AM
> To: ips@ece.cmu.edu
> Subject: iSCSI: MaxCmdSN
>
>
> In section 2.4.10, "if the PDU MaxCmdSN is less than the PDU
> ExpCmdSN ....,
> they are both ignored". Is this considered an invalid response from the
> target?. What happens when the target queuing capacity reaches 0. For
> example: target received and processed cmdsn N, but can not receive cmdsn
> N+1 yet. In the response for N, target puts N+1 for
> expectedCmdSN. How does
> it inform the initiator that its window is closed?. Setting MaxCmdSN to N
> will be ignored if I understand the above paragraph correctly.
>
> -Ayman
>
>