|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: ISCSI: CmdSN in non-leading loginJohn, My only point was that there is enough information in the section to describe correct behavior. And immediate commands do not create any blockage regardless of their CmdSN. I am still trying to understand what your point was (is). Julo
Julian, (watch out long note to follow) Sometimes our messages pass by each other in the night. I had corrected my note to indicate that I had overlooked the very obvious statement that we kept in the draft that the Login command was immediate. But the rest of the note I think is correct. That is, the command that follows the completion of the non leading-connection Login should take its CmdSN from the next command sequence number in the "Session" stream, not the CmdSN of the Login for that connection. Here is the thinking, 1. The second connection is started while commands are flowing on the Leading connection, 2. The CmdSN in the stream when the Login for the second connection is started is lets say 123. 3. Since the Login command is immediate, and does not change the ExpCmdSN then one of three things can happen a. Initiator uses the same CmdSN for the next command that it sends on the first connection, etc. (123, 124, 125, etc.) OR b. Initiator uses the next CmdSN 124 on the leading connection and thereby creates a delivery hole, for the target until the CmdSN 123 is finally received on the secondary connection and then all the commands 123----125, etc, are delivered c. No commands are sent on any connection until the Login of the secondary connection is completed and the next command is given the CmdSN of 123 4. If 3a is done, then the first command after the delivery will use the next CmdSN in the stream and no blockage would have occurred. 5. If 3b is done there will be a blockage of the delivery of commands until the secondary Login in completed 6. If 3c is done, then the first command after the delivery will use the same CmdSN (123) as the Login, but the commands will have a blockage at the initiator until the secondary Login is completed. So I believe, that IF the subsequent commands in a secondary connection has to have the same CmdSN as the Initial Login of that connection, either all commands in the entire session must stop being sent on other connections, or commands will not be delivered until the delivery hole is closed at the end of the secondary Login. If we were to handle the Login for secondary connections like other immediate commands, the 3a. path would be done (the ExpCmdSN would not be advanced, but other commands will continue and they will advance the ExpCmdSN as normal, and when the next non immediate command is sent on the same connection as the immediate command, it is simply given the very next CmdSN from the session stream. The problem of treating the secondary connection Login exactly like the leading Connection, when it comes to the next command, is that, unlike the leading connection, other things continue to move on the other connections. Therefore, in my opinion the correct way of interpreting the current wordage in 9.12.8 is at the point it says: "....and the first non-immediate command also carries the CmdSN 123.)" we should read that as "... the next non-immediate command IN THE SESSION also carries the CmdSN 123)". In this way the words mean the same for both the leading connection and secondary connection. This is how I believed the implementations have been built. Anyway, that was the essences of my words where I suggested: At the end of the first paragraph under 9.12.8, we should add, the words "Similar rules exist for the Leading Login for a non leading connection, that is, the next CmdSN value is used for the Leading Login of the non leading connection, and that same number is used with all the subsequent Login PDUs on that connection. But at completion of the Login of the connection, the ExpCmdSN is not advanced and the subsequent commands take the next CmdSN value in the session." Now perhaps we can have some better words, perhaps the words but clearly the current word have confused some folks. Perhaps even me :-) . . . John L. Hufferd Senior Technical Staff Member (STSM) IBM/SSG San Jose Ca Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688 Home Office (408) 997-6136, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com Sent by: owner-ips@ece.cmu.edu To: John Hufferd/San Jose/IBM@IBMUS
The text as it is written has not been updated since we at one point had the Login command marked as and Immediate command. Therefore, it always had to pick up the next value that would be assigned to the next non immediate command. Also since it was an immediate command the ExpCmdSN was not advanced so it did not effect the commands in the rest of the session. They would continue with the next CmdSN as usual. Login is still not suppose to advance the ExpCmdSN, so there should still not be a problem. When we removed the flag that said that it was an immediate command, the wording in the draft about the Initial Login was not adjusted. I suggest that the only thing that needs to be changed is the following: At the end of the first paragraph under 9.12.8, we should add, the words "Similar rules exist for the Leading Login for a non leading connection, that is, the next CmdSN value is used for the Leading Login of the non leading connection, and that same number is used with all the subsequent Login PDUs on that connection. But at completion of the Login of the connection, the ExpCmdSN is not advanced and the subsequent commands take the next CmdSN value in the session." This is what we were all doing, and simply did not change anything with the Immediate bit was dropped from the Login PDU. I suggest we leave it with that operational intent, since it was so long ago that we establish this way of operating I can not remember exactly why, but would not like to be surprised sometime in the future, and I do not see why we would have to change it. . . . John L. Hufferd Senior Technical Staff Member (STSM) IBM/SSG San Jose Ca Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688 Home Office (408) 997-6136, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com "Mark S. Edwards" <marke@muttsnuts.com>@ece.cmu.edu on 05/10/2002 10:38:07 AM Sent by: owner-ips@ece.cmu.edu To: John Hufferd/San Jose/IBM@IBMUS cc: <ips@ece.cmu.edu> Subject: Re: ISCSI: CmdSN in non-leading login John, If an initiator opens a new connection on a session and in the first login request uses the current session CmdSN, the lifetime of that CmdSN continues all the way through until the completion of the first non-immediate command on that new connection. This means that the initiator has to reserve the use of that CmdSN for the first non-immediate command on that particular session. That's stupid. Especially bearing in mind that it could be pumping commands like crazy down other open connections in the session. There is also another side effect, if the login takes any length of time, like computing a DH or having to make external calls to a Radius or a Kerberos server, then because this CmdSN is outstanding, the other connections in the session will block as soon as the command window is filled. The final sentence in 9.12.8 is meaningless for a non-leading connection. The target either has to block the whole session until the login is completed, or run the risk of corrupting session state. Mark At 10:14 AM 5/10/2002 -0700, John Hufferd wrote: >I do not see the issue, it works as is, it does not effect the first non >immediate command unless it completes as a authorized connection. Again, I >do not see the issue, in fact it works as is. > >. >. >. >John L. Hufferd >Senior Technical Staff Member (STSM) >IBM/SSG San Jose Ca >Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688 >Home Office (408) 997-6136, Cell: (408) 499-9702 >Internet address: hufferd@us.ibm.com > > >"Mark S. Edwards" <marke@muttsnuts.com>@ece.cmu.edu on 05/10/2002 09:47:53 >AM > >Sent by: owner-ips@ece.cmu.edu > > >To: Bill Studenmund <wrstuden@wasabisystems.com> >cc: <ips@ece.cmu.edu> >Subject: Re: ISCSI: CmdSN in non-leading login > > > >At 09:13 AM 5/10/2002 -0700, Bill Studenmund wrote: > >On Fri, 10 May 2002, Mark S. Edwards wrote: > > > > > There are a couple of problems here for the target. At what point is a > > > non-leading connection considered to be part of the session ? Is it >the > > > moment that the login request is received with a non-zero TSIH ? Or is >it > > > only when the non-leading login succeeds and it enters full feature >phase ? > > > >I think the target shouldn't let anything from a new connection influence > >things until after at least the security phase has been passed. Otherwise > >we have a security hole. After operational negotiation would probably be > >best. > >Good, as far as I can see, a new connection should not be allowed to >influence any session context until the connection reached full-feature >phase. But this is a problem if we have to follow the CmdSN rules of >9.12.8. See next comment. > > > > > So, should an initiator set the CmdSN in the first login request to >zero > > > and only synchronise with the session command stream after full feature > > > phase is established ? This is my preferred option. > > > > > > What happens if the initiator tries using the current session command > > > sequence number is that whilst the login negotiation occurs, other > > > connections within the session can be issuing new commands, so by the >time > > > that the login is finished the CmdSN exchanged in the initial request >is > > > invalid anyway. > > > > > > I would like to see something along the lines that for a non-leading > > > connection, the CmdSN field MUST be zero and that the connection can >not be > > > considered part of the session until full feature phase is entered, at > > > which point any commands issued on the connection are now synchronised >with > > > the session command sequence number as observed by all other existing > > > connections on the session. > > > >I thought login counted as a command, so it got its own command number, in > >the stream of all other commands. ?? > > >For a leading connection the CmdSN, whether zero or non-zero, is regarded >as a primer with which to set the initial session command sequence number, >all the login requests exchanged in the negotiation, no matter how many >carry the same CmdSN as does the first non-immediate command. In other >words, the first pdu in a leading connection does influence session state. > >So if we agree that a non-leading connection can not influence session >state until full-feature phase, then we have to also state that the rule in >9.12.8 about the first non-immediate command carrying the same CmdSN as the >initial login request can not work for a non-leading connection. In this >case, the first non-immediate command must be set from the next logical >value in the session context. I would like the spec to have this added and >to explicitly say that the CmdSN in a login request for a non-leading >connection is ignored by the target. I'd really prefer that the actual >value be zero, but that's not necessary from a protocol perspective if the >target ignores the field. > >Mark.
Home Last updated: Mon May 13 14:18:39 2002 10097 messages in chronological order |