RE: Re: iSCSI Marker questions

[Dean Scoville <Dean.Scoville@qlogic.com>]

Julian,

An undesirable side affect of starting the marker interval at connection establishment time and including the SYN char in the interval is that the alignment of markers will ALWAYS be different than the alignment of iSCSI PDUs. This is messy for anyone who is looking at things in 4-byte pieces. The alignment of the markers will ALWAYS differ by one byte from the PDU alignment. Using the example cited in the earlier posting, the first marker byte will ALWAYS have an EVEN-numbered TCP sequence number, and the first PDU byte will ALWAYS have an ODD-numbered TCP sequence number.

 

> Assume the following interaction:
>
> I->  SYN     (TCP sequenceNum=1000)
>
> T->  SYN-ACK (TCP sequenceNum=2000)

>
> I->  Login Request PDU, T=0,CSG=1,NSG=0:
>      InitiatorName=xxx
>      TargetName=yyy
>      SessionType=normal
>      ...
>      FMarker=send-receive
>      RFMarkInt=512,1024
>
> T->  Login Response PDU, T=0,CSG=1,NSG=0:
>      ...
>      FMarker=send-receive
>      SFMarkInt=1024
>      RFMarkInt=1024
>
> I->  Login Request PDU, T=1,CSG=1,NSG=3:
>      SFMarkInt=1024
>      (64-byte PDU... TCP sequenceNum=1301-1364)
>
> T->  Login Response PDU, T=1,CSG=1,NSG=3:
>      (48-byte PDU... TCP sequenceNum=2201-2248)
>
> The above interaction designates a 1024 x 4 = 4096-byte marker
> interval in both directions. The first PDU byte sent by the
> intitiator in full-feature mode will have sequenceNum=1365, and
> the first byte sent by the target will have sequenceNum=2249.
>
>                TCP SeqNum of    

>                First Marker    

>                ------------    

> Initiator:     5096-5103        
> Target:        6096-6103 
> 
 
Since the layer that inserts the markers needs to be aware of login anyway (it can't insert markers during login phase), I would rather that the marker interval didn't start until the end of login phase. It would avoid some messy corner cases (see next paragraph). If this is unacceptable, then let's at least not count the SYN char in the marker interval, so that the markers can be on the same 4-byte alignment as the PDUs.

 

If the marker interval remains as it is currently defined, how are partial markers treated at the end of login phase? Modifying the previous example a little, if the initiator's final PDU during login phase had TCP sequence numbers 5035-5098 should we insert the last 5-bytes of the marker into the data stream? If the final PDU were at 5039-5102, would we insert 1-byte of the marker into the data stream? Or would we skip the first marker altogether because part of it would have begun during login phase? Either way, it is messy.

 

-Dean

-----Original Message-----
From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
Sent: Friday, December 14, 2001 7:58 PM
To: Dean Scoville
Cc: ips@ece. cmu. edu (E-mail); owner-ips@ece.cmu.edu
Subject: RE: Re: iSCSI Marker questions

You are right there is no range today for SFMarkint - we may want to change this to enable a one-shot negotiation.

Julo

Dean Scoville <Dean.Scoville@qlogic.com> Sent by: owner-ips@ece.cmu.edu 14-12-01 20:10

To:        "ips@ece. cmu. edu (E-mail)" <ips@ece.cmu.edu>         cc:                 Subject:        RE: Re: iSCSI Marker questions

Julian,
Thanks for the response. Some examples would be very helpful.
Your response shows a parameter negotiation with SFMarkInt=1,512
but Draft 9 doesn't allow a range for SFMarkInt... only for RFMarkInt.
Is a range also being allowed for SFMarkInt?
 
>
> I assume a normal dialogue may go like:
>
> I->FMarker=send-receive,RFMarkInt=1,4,SFMarkInt=1,512
> T->FMarker=send-receive,RFMarkInt=8, SFMarkInt=2
>
Section 2.2.4 of the draft currently requires all integer parameter
negotiations to have a response . (Responses are optional only
with boolean negotiations where the offered value allows only a
single outcome -- such as an "OR" boolean function with a "Yes"
offer). Are responses always required for RFMarkInt and SFMarkInt,
or is the following acceptable?
 
I-> FMarker=send-receive,RFMarkInt=1,4,SFMarkInt=1,512
T-> FMarker=send,SFMarkInt=4
 
Would it also be allowed to respond with the following and, if so,
is one preferred over the other? (I'm not sure about the current
state of using "0" as a "don't care" value, but is that also
permitted in cases where the response would be irrelevant?)
 
I-> FMarker=send-receive,RFMarkInt=1,4,SFMarkInt=1,512
T-> FMarker=send,SFMarkInt=4,RFMarkInt=irrelevant
 
With the following exchange, is the target required to respond
with any RFMarkInt or SFMarkInt values, since there is no
real choice of values? Agreeing to send markers implies that
the interval is 512, and not agreeing to receive markers implies
that there is no receive interval. Is the following allowed, or
should "RFMarkInt=4" and "SFMarkInt=irrelevant" be returned
by the target?
 
I-> FMarker=send-receive,RFMarkInt=4,SFMarkInt=512
T-> FMarker=send
 
Is the following acceptable, even though it doesn't reply to
RFMarkInt and SFMarkInt?
 
I-> FMarker=send-receive,RFMarkInt=4,SFMarkInt=512
T-> FMarker=no
 
It would also be helpful to explain what happens if the login
phase goes beyond the negotiated first marker interval.
Assume, for example, the following:
 
  SYN - at TCP sequence num 1000.
  SYN-ACK - at TCP sequence num 2000.
  negotiate to FMarker=send-receive,RFMarkInt=100, SFMarkInt=100.
 
If the 1st byte of 1st initiator PDU in full-feature mode has TCP seqNum=1404,
full-feature phase starts in the middle of what would have been a marker.
Is a partial marker inserted? What would be the TCP seqNum of the first and
second marker?
 
If the 1st byte of 1st target PDU in full-feature mode has TCP seqNum=2810,
what would be the TCP seqNum of the first and second marker?
 
thanks,
Dean
-----Original Message-----
From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
Sent: Tuesday, December 11, 2001 4:54 PM
To: ips@ece.cmu.edu
Subject: Fw: Re: iSCSI Marker questions


----- Forwarded by Julian Satran/Haifa/IBM on 12-12-01 02:39 -----

Julian Satran 11-12-01 14:44

To:        IPS List        cc:                Subject:        Re: iSCSI Marker questionsLink

Dean,



owner-ips@ece.cmu.edu wrote on 11-12-2001 03:09:11:

> The iSCSI Draft 9 Appendix C makes the following statements about
> Markers and the Initial Marker-less Interval:
>
>      "The offset to the next iSCSI PDU header is counted in terms
>       of the TCP stream data. Anything counted in the TCP
>       sequence-number is counted for the offset. Specifically this
>       includes any bytes "inserted" in the TCP stream by an UFL and
>       it excludes any other markers inserted between the one we are
>       examining and the next PDU header."...
>
>      "To enable the connection setup including the login phase
>       negotiation, marking (if any) is started only at the first
>       marker interval after the end of the login phase."
>
> I understand that markers are not inserted until after login phase.
> Am I correct to assume that the placement of the first marker
> determined by the TCP sequence numbers on the final Login Request/
> Response PDUs, or is initial marker position determined by the
> TCP sequence numbers at connection establishment?
>
> Assume the following interaction:
>
> I->  SYN     (TCP sequenceNum=1000)  -- irrelevant to this discussion?
>
> T->  SYN-ACK (TCP sequenceNum=2000)  -- irrelevant to this discussion?
>
> I->  Login Request PDU, T=0,CSG=1,NSG=0:
>      InitiatorName=xxx
>      TargetName=yyy
>      SessionType=normal
>      ...
>      FMarker=send-receive
>      RFMarkInt=512,1024
>
> T->  Login Response PDU, T=0,CSG=1,NSG=0:
>      ...
>      FMarker=send-receive
>      SFMarkInt=1024
>      RFMarkInt=1024
>
> I->  Login Request PDU, T=1,CSG=1,NSG=3:
>      SFMarkInt=1024
>      (64-byte PDU... TCP sequenceNum=1301-1364)
>
> T->  Login Response PDU, T=1,CSG=1,NSG=3:
>      (48-byte PDU... TCP sequenceNum=2201-2248)
>
> The above interaction designates a 1024 x 4 = 4096-byte marker
> interval in both directions. The first PDU byte sent by the
> intitiator in full-feature mode will have sequenceNum=1365, and
> the first byte sent by the target will have sequenceNum=2249.
>
> Assuming the markerless interval starts at the end of login
> phase, the first two markers in each direction will have the
> following TCP sequence numbers:
>
>                TCP SeqNum of    TCP SeqNum of
>                First Marker     Second Marker
>                ------------     -------------
> Initiator:     5461-5468        9565-9572
> Target:        6345-6352        10449-10456
>
No - the correct numbers are dependent only on the marker interval (not the length of the login phase) and are:

Initiator        5096-5103        9200-9201
Target           6096-6103        10200-10201
 

> Is this the correct interpretation of marker usage in iSCSI
> Draft 9, or does marker placement depend on the connection's
> initial sequence numbers?
>
> Also, is "RFMarkInt=..." always considered an offer, and "SFMarkInt="
> considered a reply to that offer? If an offer is sent with "FMarker=..."
> and "RFMarkInt=...", MUST the reply contain either "FMarker=no" or
> BOTH "FMarker=yes" and "SFMarkInt=..."?
>

Fmarker is not boolean - legal values are no, send, receive, send-receive
The sender and receiver must set the interval it wants/is ready to use
otherwise the responder can't answer.
I assume a normal dialogue may go like:

I->FMarker=send-receive,RFMarkInt=1,4,SFMarkInt=1,512
T->FMarker=send-receive,RFMarkInt=8, SFMarkInt=2

Please observe that target answers with RFMarkInt to the initiators SFMarkInt and viceversa.

I will attempt an example in draft 10 (last?).



> Thanks,
> Dean Scoville
> QLogic Corp.