Re: [Idr] AD Review of draft-ietf-idr-bgp-gr-notification-13

["John G. Scudder" <[e-mail]>]

(with co-author hat)

Hi Alvaro,

> On Dec 7, 2017, at 4:11 PM, Alvaro Retana <[e-mail]> wrote:
> 
> Dear authors:
> 
> I just finished reading this document — I have some comments, please see the list below.
> 
> I understand the intent of this document: instead of resetting a BGP session when a NOTIFICATION is received, use Graceful Restart; if the session is going to *really* be reset, then use the new Hard Reset sub-code.  That makes sense to me…but, is that the only code/sub-code for which it makes sense to do a hard reset?  The NOTIFICATION has always had the “stigma” of being something bad, so much that we (idr/IETF) have even worked on ways to reduce its use (rfc7606, for example).  I want to ask the WG to consider whether other code/sub-code NOTIFICATION combinations should also result in a hard reset.  I think there are several cases, for example:
> 
> (1) rfc4486 (Subcodes for BGP Cease Notification Message) defines “Administrative Shutdown” ("a BGP speaker decides to administratively shut down its peering with a neighbor”).  It seems to me that the sender of this NOTIFICATION would not want to "follow the rules for the Receiving Speaker” (as specified in Section 4).
> 
> (2) rfc4486 also defines "Administrative Reset” ("a BGP speaker decides to administratively reset the peering with a neighbor”); no more details are provided, but that sounds like a hard reset to me.
> 
> (3) …there’s probably others...
> 
> Having said all that, I note that Section 3.1. (Sending a Hard Reset) specifies the “encapsulation” (for lack of a better word) of the real reason for the Hard Reset.  If the consensus is to go forward with that, and not call out other exceptions, then I think that the text in 3.1 should expand more on the encapsulation operation and the rationale for doing it this way, and the document should also address other recent work that recommends the use of Administrative Shutdown, for example draft-ietf-grow-bgp-session-culling (a BCP currently in the RFC Editor’s Queue).
> 
> [After I wrote the text above…] I found that some of the points have been discussed on the list already — please include some of that discussion/analysis in the document.

I'll draft something and we can discuss.

Just FYI, the encoding in question was adopted in version -03 as a result of Chris Hall's comments during the WGLC way back in the day (https://www.ietf.org/mail-archive/web/idr/current/msg13205.html)

> I’ll wait until the issue above and ones marked Major (below) are addressed before starting the IETF LC.
> 
> Thanks!
> 
> Alvaro.
> 
> 
> 
> Major:
> 
> M1. Unfortunately, rfc4724 failed to setup a registry for the Restat Flags (or the Flags for Address Family), which means that anyone is able to use the bits in there (assuming the receiver looks at them, of course).  Given that there are only a few bits, and to prevent conflicts, I would really like to see a registry set up.  This document is already tagged to Update rfc4724, so it seems like a good place to establish the registries.  [If for some reason you rather not include that information here, then we can take care of it elsewhere.  IOW, this request is not a requirement.]

I'll do this and while I'm at it, set one up for Flags for Address Family too.

> M2. Section 4.1. (Rules for the Receiving Speaker) has me a little confused.  Are the proposed changes contingent to setting the N bit?  

No.

> The text starts by saying: "As part of this extension, routes from the peer previously marked as stale MUST NOT be deleted, until and unless the optional timer…expires…”…does that mean that the timer is no longer optional?   

No, it's just that we foolishly didn't give the timer a name in the base GR spec. Here's the paragraph from RFC 4272:

   To put an upper bound on the amount of time a router retains the
   stale routes, an implementation MAY support a (configurable) timer
   that imposes this upper bound.

In practice, all implementations have this and call it something like 'stale timer' or whatever. The lack of a defined name for this thing makes it a pain to write about it, which I think is what you've bumped into here. I'll propose a replacement in OLD/NEW style.

> Then you also say: “...if the Graceful Notification ("N") bit is not set in the newly received Graceful Restart Capability, no new actions are triggered on the Receiving Speaker -- in particular, a clear "N” bit does not trigger deletion of stale routes.”  If I understand rfc4724 correctly, stale routes could be deleted — the text indicates changes in the behavior even if the N bit is not set, right?  If you are Updating this section of rfc4724, what would make it crystal clear is an “OLD/NEW” notation of the text (as in, this is the OLD text…and this is the NEW text…).

Oh ouch, I see why you were confused. I believe we meant F bit, not N bit. I'll propose a replacement in OLD/NEW style.

> M3. Security Considerations:  Maybe not a security issue, but something to think about.  Section 4.1 says that “routes...previously marked as stale MUST NOT be deleted, until and unless the optional timer...expires, or unless a Hard Reset is performed.  This supersedes the “consecutive restarts” requirement…”.  Not deleting the stale routes and not making the timer mandatory could result in stale routes that live forever if an attacker manages to create consecutive restarts (by simply sending NOTIFICATIONS before EoR) — stale routes are ok in the short term, but may point in the wrong direction eventually.  Is this an issue?  I think that it would be mitigated if the timer was made mandatory (with a nice default).

That seems like a good idea. While we're at it, we can give it a name. 

> Minor:
> 
> P1. Section 4: “...receive and send BGP NOTIFICATION messages in Graceful mode...” What is “Graceful mode”?

Changed to "according to the procedures of this document".

> Nits:
> 
> N1. In Section 2, please indicate that the first figure corresponds to the GR Capability from rfc4724.

Actually unless someone objects I'll remove the figure entirely. I'm not sure why we put it in. It's not like this document stands alone anyway.

> N2. s/subcode is defined known as/subcode is defined as

OK

> N3. s/Graceful Notification flag/Graceful Notification bit   (For consistency)

OK

> N4. The operation is obviously per-AF; maybe it’s worth saying that somewhere just for completeness.

I don't think I agree -- the extension is NOT per-AF, it applies to all AFs for which GR is enabled (that's why the N flag goes in the Restart Flags field and not the Flags for Address Family field). The underlying operation of GR continues to be per-AF, it's true, but that's not affected by this spec so I think it might be confusing to add words about it?

Thanks,

--John