Re: [Gen-art] Gen-art last call review of draft-ietf-pwe3-cbit-negotiation-04

[Elwyn Davies <[e-mail]>]

Hi, Lizhong.

I have removed the bits that we are agreed on to shorten the mail a bit.

I don't think we have converged just yet.

Regards,
Elwyn

On Wed, 2012-06-20 at 01:17 +0800, Lizhong Jin wrote:
> 
> Hi Elwyn, 
> Thank you for the prompt reply. See inline below. 
> 
> Lizhong 
>   
> 
> Elwyn Davies <[e-mail]> wrote 2012/06/19 23:16:08:
> > > > Document: draft-ietf-pwe3-cbit-negotiation-04.txt
> > > > Reviewer: Elwyn Davies
> > > > Review Date: 19 June 2012
> > > > IETF LC End Date: 15 June 2012
> > > > IESG Telechat date: 21 June 2012
> > > > 
> > > > Summary: 
> > > > Not ready. The proposal for the NOT PREFERRED to PREFERRED
> > > transition
> > > > case does not appear to be compatible with the existing RFC 4447
> > > > standard in the way stated and there are a number of other minor
> > > issues.
> > > > The draft is also in need of an editing pass by an author whose
> > > mother
> > > > tongue is English as there are parts where the syntax is misleading
> > > > Major issues: 
> > > > s3: The discussion of the NOT PREFERRED to PREFERRED transition case
> > > > (buried at the end of s3 - causing me to ask 'what about this case?'
> > > > during reading of s2 and most of s3)  implies that some existing RFC
> > > > 4447 procedure applies. Clearly, if the PW is not using the control
> > > word
> > > > then there is nothing to do.  On the other hand, inspection of s6.2
> > > of
> > > > RFC4447 indicates that once the two PEs have agreed on c = 1, 'setup
> > > is
> > > > complete' and Label Mapping messages would therefore be
> > > > 'unexpected' (see item '-i' in second set of bullets in s6.2 of RFC
> > > > 4447). So, what procedure is to be used? And what implications does
> > > this
> > > > have for backwards compatibility?  Wouldn't it be generally simpler
> > > to
> > > > apply the PREFERRED to NOT PREFERRED mechanism to all case? 
> > > [Lizhong] this draft is to solve the case to change a PW from c=0 to
> > > c=1, that means one PE should change its use of control word from NOT
> > > PREFERRED to PREFERRED. RFC4447 is already there and deployed, and the
> > > PE will not always send its locally configured preferrence according
> > > RFC4447, see PE1 behavior in section 2 step 1&2. That's why we could
> > > not simply apply the mechanism at the end of section 3. Hope it is
> > > clear. 
> > There are two issues here:
> > - Clarity: RFC 4447 does not have any discussion of what might happen if
> > the configuration value changes.  The draft focusses on on transition
> > direction but does not mention the other except for one paragaph right
> > at the end of s3.  It is therefore reasonable that somebody looking at
> > this draft would wonder 'what about the other transition direction?'.
> > Whether or not anything needs to be done it would save people wondering
> > if you put in a sentence in the introduction to explain that the other
> > direction is not (or is) a problem. 
> [Lizhong] accept. Add following: 
> When PE changes the preference for the use of control word from
> PREFERRED to NOT PREFERRED, it should follow [RFC4447], and there is
> no problem. 

See below...
> 
> > - Technical: The paragraph in s3 implies that the PREFERRED to NOT
> > PREFERRED direction will result in some part of the RFC 4447 protocol
> > being (re-)invoked.  What part is not made very clear.  As far as I can
> > see sending more messages other than Label Release or Label Withdraw for
> > this PW is not part of the RFC 4447 protocol and hence will cause an
> > error.  Alternatively if it isn't reinvoked, the PW will continue to use
> > control words.  Please explain what is going on here.  The fact that RFC
> > 4447 has been deployed doesn't explain what is going on. 
> [Lizhong] now I understand you concern. The last sentance of s3 is not
> clear. How about the following: 
> In that case, local PE will always send Label Withdraw message if
> already sending Label Mapping message, and then send new Label Mapping
> message with C-bit value following the procedures defined in
> [RFC4447]. 

I think we have reached the heart of the issue here:  RFC 4447 does not
mention changing the control word configuration.  The combination of the
existing text in the draft and what you have just weittrn above
indicates that *in both transition directions* the PE changing
configuration notifies its peer by sending a Label Withdraw. It is not
immediately obvious to me that you could deduce this from RFC 4447 - if
it is then there needs to be a pointer to the relevant section in RFC
4447 to enlighten the ignorant like me.  Otherwise the PREFERRED to NOT
PREFERRED transition is *not* just 'follow the RFC 4447 procedures'.

If I now understand correctly, the situation is as follows:
- in both cases when there is a control word preference transition, send
a Label Withdraw.
- in the NOT PREFERRED to PREFERRED case send Label Release also (order
not important), wait for a responses from peer, then send Label Request
and negotiate use of control word using Label Mapping C bit as for a new
PW (might result in either using or not using control word depending on
state in peer).
- in the PREFERRED to NOT PREFERRED case, just negotiate C bit to 0
using Label Mapping - label will be maintained. [Need to check that RFC
4447 is expecting/will allow this sequence.]

If I have this straight, I think that the PREFERRED to NOT PREFERRED
case needs some more words in s3 and an introduction in s2. (And it
would be easier to understand with this case in a separate sub-section -
sorry to harp on about this).   

> > > 
> > > > 
> > > > Minor issues: 
> > > > s3: Has there been any discussion on possible race conditions?
> > >  Changing
> > > > the configuration value during the message exchange strikes me as
> > > > dangerous - it is probably sufficient to note that changes should be
> > > > suppressed during the Label Mapping message exchange but I am not
> > > > totally sure about this. 
> > > [Lizhong] The message would be processed in sequence in TCP-based LDP
> > > session. I do not see a problem here. But we do have a note in section
> > > 3 for multi-segment PW for sequence processing. 
> > There are several network round trips in the message exchanges.  The
> > protocol and the configuration mechanism in a PE are potentially
> > separate threads.  There is scope, depending on the implementation, for
> > the user at the 'remote' PE to change the configuration during the
> > message exchange making for a potential race condition. 
> [Lizhong] we discussed the race condition on the PWE3 maillist from
> Spike, and for multi-segment PW, we add a note to ensure the sequence
> processing for implementation.
I look forward to the revised text.
>  
> 
> > > 
> > > > 
> > > > s3, bullet '-i':  I completely misparsed this section on first
> > > reading
> > > > and I am still not absolutely sure what message sequence is being
> > > > specified.  Working back from later sections I *think* that the
> > > > intention is:
> > > > IF Mapping sent THEN { send Withdraw; send Release;}
> > > > Wait to receive Release
> > > > The implication at present is that a Mapping might not have been
> > > sent
> > > > and then only a Release is needed: is this a possibility? Please
> > > > clarify.
> > > > The picture in Appendix A suffers from the same problem. 
> > > [Lizhong] Adrian raise the same comment, and I change it as below: 
> > >        -i.  PE MUST send a Label Release message to remote PE. If a
> > > PE 
> > >             has previously sent a Label Mapping message to a remote
> > > PE, 
> > >             it MUST also send a Label Withdraw message to the remote
> > > PE, 
> > >             and wait until it receives a Label Release message from
> > > the
> > >            remote PE. 
> > What does it send first if both must be sent?  The text in the rest of
> > the document implies withdraw then release.  This new text sort of
> > implies release then withdraw but isn't really clear. 
> [Lizhong] both should be sent, and does not require the sending
> sequence. The two messages does not have dependence. 
In that case, it is important to say so.
> 
> > > 
> > > > 
> > > > s3, discussion of multi-segment PWs: The statement that S-PE's
> > > SHOULD
> > > > assume an initial passive role seems to have several problems:
> > > > - Does this mean that changing the configuration of an S-PE would
> > > not
> > > > provoke the new mechanisms?
> > > > - The passive role situation is only specified for some sorts of
> > > linked
> > > > FECs in RFC 6073 - what about other cases?
> > > > - What are the consequences for ignoring the SHOULD in this case? (I
> > > > have to say I am unsure that RFC 6073 deals with this problem
> > > either.) 
> > > [Lizhong] we follow RFC6073, and passive role is only applied for the
> > > PW FEC, other cases are out of scope. 
> > Why? This needs an explanation. Also this doesn't cover the point about
> > whether reconfig of an S-PE is allowed and how this squares with the
> > passive role.
> > > SHOULD means highly recommended. We do not meet any problem when
> > > implementing RFC6073. Do you suggest to change this to MUST? 
> > Whenever a specification includes a 'SHOULD' the question arises of what
> > alternatives there might be and what is the reason for preferring the
> > suggested solution.  In the original setup, it is fairly clear that life
> > is easier letting setup progress from  one end to the other (although
> > this is not spelt out - I would have argued for this had  I reviewed the
> > doc).  For the configuration change this is less obvious.  Regarding the
> > point about reconfig of the  S-PE, it is going to make life more
> > complicated if the S-PE has to tell the T-PE to be active so it can be
> > passive.  I would therefore argue that probably the notion of passivity
> > is irrelevant to the transition use case.  Whether it should be SHOULD
> > or MUST is therefore moot. 
> [Lizhong] how about the following? Because we just refer to RFC6073, and does not add anything new. 
> An initial passive role is defined in [RFC6073] for S-PE.
I can't see that this really helps.
There are still two points at issue here:
- Why are the cases other than the one in RFC 4447 that refers to a
passive S-PE out of scope?  There aren't any statements to this effect
in the document.
- It is difficult for a S-PE that has its control word configuration
change to act passively.  If it is really passive than nothing will
happen until the label is withdrawn.  Otherwise it can't be descibed as
passive for this action.  The text in the current draft actually
describes what is specified in RFC 6073 when talking 'passive' rather
than explaining what happens when the control word configuration
changes.
>  
> > > [Lizhong] has been revised in v-04.
> > Not as far as I can see. 
> [Lizhong] write "pseudowire" directly in the text, not PW. 
There are lots of instances of PW still left in v04. Expand acronym =>
define what it means on first usage i.e pseudowire (PW).  Same for the
various other acronyms that aren't defined in the text.
> 
> > > > 
> > > > s2: Expand acronym PE. 
> > > [Lizhong] accept 
> > > 
> > > > s2, 2nd sentence: s/configurable/configured/ 
> > > [Lizhong] configurable is right. 
> > Leave that one to the RFc Editor.
> > > 
> > > > s2, 3rd and 4th sentences: I *think* this text is trying to say: 
> > > >   The intention of the control word negotiation is that the control
> > > word
> > > > will be used when both endpoints are configured with control word
> > > usage
> > > > PREFERRED.  However if one endpoint is initially configured with
> > > control
> > > > word usage NOT PREFERRED but later changes to PREFERRED, a PW
> > > between
> > > > the endpoints will not transition to usage of the control word as
> > > > explained below. 
> > > [Lizhong] no, the case is that operator deploys PW with control word
> > > used in the first phase. In the second phase, they want to upgrade
> > > their PW service to use control word. Two different deployment
> > > timeframes.
> > That is what my sentence says. 
> [Lizhong] ok. 
> 

> > > > s3: It would be much clearer if s3 was divided into 3 sub-sections
> > > > (possibly reordered):
> > > > - PREFERRED to NOT PREFERRED transition
> > > > - NOT PREFERRED to PREFERRED transition
> > > > - Multi-segment case (which should refer to both previous cases)
> > > > The pointer to the diagram in Appendix A could usefully occur in the
> > > > introduction to s3.  If this was adopted s3.1 could either be a
> > > fourth
> > > > sub-section or a sub-sub-section of the PREFERRED to NOT PREFERRED
> > > > section. 
> > > [Lizhong] PREFERRED to NOT PREFERRED does not introduce any new.
> > > Multi-segment case is fully inherited from single-segment case. It
> > > would be redundancy to have sub-sections here. 
> > I disagree strongly.  The multi-segment case affects the other RFC and
> > needs to be made to stand out so that implementors can see where the
> > changes occur.  The PREFERRED to NOT PREFERRED case also ought to be
> > separated whether or not I am right about this case. 
> [Lizhong] how about only divide multi-segment? The case of PREFERRED to 
> NOT PREFERRED case is only for reference, and the text is not many.
That is so, but it is a change of subject and the fact that it will then
be in the table of contents makes it easier for implementors to find the
relevant text.  I wouldn't call it 'redundant'.

> > > > s3, last para/Appendix A:  The diagram doesn't cover the PREFERRED
> > > to
> > > > NOT PREFERRED transition. 
> > > [Lizhong] no, there is a "no" branch under "NOT Preferred to
> > > Preferred" to cover "PREFERRED to NOT PREFERRED". Anyway, the diagram
> > > discription should be "NOT Preferred to Preferred" which is wrong in
> > > v-04. 
Need to check it reflects the Label Withdraw message needed in the
PREFERRED to NOT PREFERRED case.