Re: [Technical Errata Reported] RFC5880 (7240)

[Reshad Rahman <[e-mail]>]

 Hi John,
I agree with Jeff's comments.
Regards,Reshad.
    On Tuesday, March 7, 2023, 12:57:28 PM EST, Jeffrey Haas <[e-mail]> wrote:  
 
 John,

The text below covers what I'd hoped to accomplish with the errata.  Thanks.

At some point BFD should consider if we feel like doing the work to move the base documents to Internet Standard.  

-- Jeff


> On Mar 6, 2023, at 9:42 AM, John Scudder <[e-mail]> wrote:
> 
> Once again, we seem to have come close to something resembling a consensus on this point, other than needing some new text to agree on. To spur discussion, I’ve updated the erratum, but not verified it, pending discussion of my suggested text. I’d verify it as “hold for document update”.
> 
> Please weigh in, within the next few days, if you want to express any further opinions. Either “good grief, end the pain and verify it already” or “no John that’s wrong, please change it as follows” would be helpful.
> 
> I’ve pasted the updated version below.
> 
> —John
> 
> 
> Section 6.8.1 says:
> 
>  bfd.LocalDiag
> 
>      The diagnostic code specifying the reason for the most recent
>      change in the local session state.  This MUST be initialized to
>      zero (No Diagnostic).
> 
> It should say:
> 
> No proposed changes are offered here. See the notes for further discussion.
> 
> Notes:
> 
> RFC 5880 at various points calls out setting the value of bfd.LocalDiag as part of state transitions. The text defining the feature calls for it to be initialized to zero. Discussion on the WG mailing list following the filing of the initial version of this erratum revealed two things:
> 
> First, the text of the RFC is correct, complete, and reflects the authors’ intention at the time of writing, which really WAS that the value should only be initialized to zero but not reset to zero at any other time. 
> 
> Second, by not emphasizing this point, the spec although formally speaking unambiguous, left space for implementors to exercise their intuitions and creativity. As a result, several implementations are reported to reset this value to zero when the session transitions back to Up.
> 
> The discussion is archived at https://mailarchive.ietf.org/arch/msg/rtg-bfd/yEOx2LTO51zq1he6vChUOVJySqM/ . If a new version of RFC 5880 is prepared in the future, this question should be reopened as part of that process. It would also be possible to offer a standards track document to update RFC 5880 in this respect if WG consensus can be found for a new approach. 
> 
> 
>> On Feb 22, 2023, at 5:36 PM, Dave Katz <[e-mail]> wrote:
>> 
>> Cool.  And I refer once again to the paragraph in the spec about pedantic coders.  If the WG actually approves anything that reads the value of the diag field, I guess they can address all of the anomalies and hope for the best…
>> 
>> —Dave
>> 
>>> On Feb 22, 2023, at 2:30 PM, Jeffrey Haas <[e-mail]> wrote:
>>> 
>>> [External Email. Be cautious of content]
>>> 
>>> 
>>> "Hold for update" was the expected outcome for my filing of the errata.
>>> 
>>> At best, we're telling new implementors that there's an issue here of note in the protocol.  The mail discussion will note that there are multiple existing implementations that have historically set the value to 0 when transitioning to Up.
>>> 
>>> It'll also log some of what your thinking was at the time.  Since RFC 5880 already talks about cases where the value is of interest (concatenated path, etc.) implementors already know to pay attention to the value even when state transitions aren't happening.
>>> 
>>> Part of my own motivation to have the behavior clear has been other proposals we've seen come and go trying to use Diag as a much more rigorous mechanism to trigger behaviors.  I had thought I could find a draft I saw during the mpls session at IETF 115 along these lines, but I appear to be mistaken.  In any case, people keep wanting to use Diag for Clever Things and it'll bite them in unpleasant places if they do so.
>>> 
>>> Greg may have recollection of the proposal I'm thinking about.
>>> 
>>> -- Jeff
>>> 
>>> 
>>> 
>>> 
>>>> On Feb 22, 2023, at 2:34 PM, Dave Katz <[e-mail]> wrote:
>>>> 
>>>> Thanks for the background.
>>>> 
>>>> I guess the fact of the matter is that, since the issue cannot affect interoperability, it’s hard to imagine getting the WG to go through a bunch of machinations to go out of its way to fix something that is entirely pedantic.  In that case I think holding the erratum for update is the right choice.  The erratum can describe the ambiguity and the WG can decide what to do about it if they find another reason to update the spec...
>>>> 
>>>> —Dave
>>>> 
>>>> 
>>>>> On Feb 22, 2023, at 11:29 AM, John Scudder <[e-mail]> wrote:
>>>>> 
>>>>> Hi All,
>>>>> 
>>>>> Regarding Jeff’s "Given the maturity of the feature, I'd suggest sticking to the reality on the ground”, I want to step in and remind folks about what our guidelines are for processing errata [1]. They are fairly narrow, by design. One of the high-order bits of the guidelines is, "Errata are meant to fix "bugs" in the specification and should not be used to change what the community meant when it approved the RFC.”  What I take this to mean in the current context is, if the behavior specified in the RFC has been found to be ‘wrong’ (for whatever definition of ‘wrong’ you choose to apply), an erratum is definitively not the way to correct that. An erratum is to clarify or correct whatever the intent of the RFC was at time of publication. Of course, many RFCs (this one included, it seems) didn’t receive detailed scrutiny of every crevice of the spec before being declared ready for publication, and so it’s not always possible to really say that the consensus was firmly one way or the other. In such cases, I think we have to err on the side of what the words in the spec say.
>>>>> 
>>>>> About the furthest I’d go in documenting that (part of) the WG now thinks the specified behavior is undesirable, is to note it in a ‘hold for document update’ erratum. That seems reasonable in this case — it can lay out the pros and cons and at least creates an artifact for future implementors to notice.
>>>>> 
>>>>> The bottom line is that changes to specified behavior require WG and IETF consensus, and that means they require an RFC to update or obsolete the old behavior. This is one of the pointy bits of our process, that RFCs document the consensus at a moment in time, not the evolving consensus.
>>>>> 
>>>>> Thanks,
>>>>> 
>>>>> —John
>>>>> 
>>>>> [1] https://urldefense.com/v3/__https://www.ietf.org/about/groups/iesg/statements/processing-errata-ietf-stream/__;!!NEt6yMaO-gk!CSN9QP45mGgvK5UVJPyirmK_RjKBRT3v8KZr_KsbVx5QvoBWSZ_k2lXd4YgZchaxL6ItLKD8ee1J$
>>>>> 
>>>>>> On Feb 22, 2023, at 11:14 AM, Jeffrey Haas <[e-mail]> wrote:
>>>>>> 
>>>>>> 
>>>>>> Dave,
>>>>>> 
>>>>>> Just as a reminder, the context for why this errata is being discussed is this inquiry:
>>>>>> https://urldefense.com/v3/__https://mailarchive.ietf.org/arch/msg/rtg-bfd/YIeCo-nQicI_OIcVncYaJM5Zz6c/__;!!NEt6yMaO-gk!CSN9QP45mGgvK5UVJPyirmK_RjKBRT3v8KZr_KsbVx5QvoBWSZ_k2lXd4YgZchaxL6ItLDNVBYRz$
>>>>>> 
>>>>>> More below:
>>>>>> 
>>>>>> 
>>>>>>> On Feb 17, 2023, at 12:04 PM, Dave Katz <dkatz=[e-mail]> wrote:
>>>>>>>> On Feb 17, 2023, at 8:47 AM, Reshad Rahman <[e-mail]> wrote:
>>>>>>>> Having the diag field as breadcrumb has been extremely useful indeed. But both ends can store last diag field sent/received, we don't have to keep sending the diag field after the failure has cleared. It seems odd to be sending a diag field which happened e.g. a year ago.
>>>>>>> 
>>>>>>> That property helped me when debugging my implementation, as it survives the restart/reboot of the far end.
>>>>>>> 
>>>>>>> There is also no timeout that would make sense;  “forever, for small values of ‘forever'” is semantically consistent and the only thing that makes sense (to me, at least).
>>>>>>> 
>>>>>>> Resetting it to zero only deletes information (albeit a tiny amount of it) and doesn’t help anything;  you know that the session is up, so the diagnostic for its most recent transition to non-upness is disambiguated.
>>>>>>> 
>>>>>>> Debugging broken things is a scramble for bits of data;  leaving a breadcrumb is a polite gift.
>>>>>> 
>>>>>> From my perspective, the breadcrumb is useful to note during the transitions, and not simply the transitions for the state.  Examples have been given where the diag is updated as part of a state transition (governed by normative text in 5880), or transitions that may happen while the session remains up (e.g. concatenated path down, echo, etc.).  The RFC isn't great about saying how you clear such things when the state is still Up; intuitively it's to return to "No Diagnostic".
>>>>>> 
>>>>>> However, your own leaning, Dave, is "leave it set forever".  Using the above examples for diag signaling an event while leaving the state up, I don't think you mean that.
>>>>>> 
>>>>>> So, again, the interesting breadcrumbs are when Things Change.  Each of these items is an edge transition of note. If I care about the event, I care about it when it happens and will remember it.  I'm not going to look at diag to reflect this forever.
>>>>>> 
>>>>>>> 
>>>>>>>> 
>>>>>>>> Also the text in 6.8.1 says "The diagnostic code specifying the reason for the most recent change in the local session state.". To me that means resetting bfd.LocalDiag to 0 when the state changes to Up.
>>>>>>> 
>>>>>>> Thus the language that needs fixing (I know, I wrote it...)
>>>>>>> 
>>>>>>>> 
>>>>>>>> AFAIK IOS-XR and JunOS reset LocalDiag. It'd be good to hear from other implementations.
>>>>>>> 
>>>>>>> Probably.  I might have even coded it that way 20 years ago, or someone else did later, thus underscoring the largely-irrelevant nature of this discussion...
>>>>>> 
>>>>>> I did confirm with Juniper's BFD developers that it's reset to 0 when we transition to Up.
>>>>>> 
>>>>>> Given the maturity of the feature, I'd suggest sticking to the reality on the ground.
>>>>>> 
>>>>>> -- Jeff
>> 
>