|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: iSCSI: update on OOO CmdSNs/connectionJulian, I am surprised to see the text in the latest rev of the doc change as suggested by Mallikarjun. I did not think there was a consensus on this subject. Just because I did not respond to Mallikarjun's last comment publicly should not be construed as agreement. Having the last word is hopefully not assumed to be consensus, otherwise a thread may never end. Somesh > -----Original Message----- > From: owner-ips@ece.cmu.edu [mailto:owner-ips@ece.cmu.edu]On Behalf Of > Somesh Gupta > Sent: Friday, November 09, 2001 3:26 PM > To: ips > Subject: RE: iSCSI: update on OOO CmdSNs/connection > > > Mallikarjun, > > I never liked the SHOULD. It is not a design point. > If we really want to allow it, perhaps a negotiation > parameter is a better choice (which is only > marginally better). Error detection and recovery > have completely different design requirements than > the data path. > > So ideally we say MUST or nothing > or we negotiate it > > Also on your point A, it "MUST" be a MUST on > a single connection case except for when required > for error recovery (i.e. the very very rare - > case of digest error). > > Somesh > > > -----Original Message----- > > From: owner-ips@ece.cmu.edu [mailto:owner-ips@ece.cmu.edu]On Behalf Of > > Mallikarjun C. > > Sent: Friday, November 09, 2001 1:49 PM > > To: ips > > Subject: iSCSI: update on OOO CmdSNs/connection > > > > > > To those interested in this discussion: > > > > Julian and I had a phone conversation on the topic > > of OOO CmdSNs on a connection. An update follows. > > > > Julian agrees that there are valid error recovery > > scenarios where CmdSNs will legitimately end up OOO > > on a given connection. > > > > OTOH, I agree with two of Julian's scenarios that he > > pointed out right away - the "cleaning command" (command > > required to be sent after a retry copy to ensure flushing > > within 2^31 -1), and an immediate Logout posted with > > unacknowledged commands. Neither of this can be shipped > > OOO - since the former undoes the flushing intent, and > > the latter breaks the rule that nothing more follows a > > Logout on the connection (and troublesome in other ways, > > see below). > > > > In general, I share the concern with Julian that we > > have not closely scrutinized all possibilities. > > > > With that said, something along the following lines > > seemed reasonable - > > > > A)Initiator MUST send commands in increasing order of > > CmdSN on a connection if both the following are true - > > - operational ErrorRecoveryLevel is 0, > > - MaxConnections is negotiated to 1. > > B)In all the other cases, initiator SHOULD send commands > > in increasing order of CmdSN on a connection. It is > > strongly encouraged that commands with out-of-order > > CmdSNs be sent on a connection only if they are > > retransmitted commands due to digest error recovery > > and connection recovery. > > > > I also suggest the following upon further reflection- > > > > C)Add wording in section 2.2.2.1 to mandate that > > the cleaning command MUST be sent in-order after > > the retried command. > > D)Warn clearly that sending an immediate Logout command > > in the presence of other unacknowledged commands MAY > > create inadvertent discarding of certain commands (even > > if it is a recovery Logout), and MAY cause protocol > > errors leading to ungraceful shutdown of the connection. > > > > Hopefully A will bring the determinism that Somesh was > > looking for certain design points. B describes the more > > general n-connection session case. C & D are fixes for > > two identfied areas (so far) which will break. > > > > Comments? > > -- > > Mallikarjun > > > > > > Mallikarjun Chadalapaka > > Networked Storage Architecture > > Network Storage Solutions Organization > > MS 5668 Hewlett-Packard, Roseville. > > cbm@rose.hp.com > > >
Home Last updated: Wed Nov 14 09:17:44 2001 7810 messages in chronological order |