On Mon, 5 Mar 2018, Ron Bonica wrote:


Please review draft-bonica-intarea-frag-fragile-01 and provide comments. The URL is https://tools.ietf.org/html/draft-bonica-intarea-frag-fragile-01.

I like it.

4.6. There are cases where this "misconfiguration" is due to vendor default not being changed. I do not equate "misconfiguration" with "didn't change default configuration". Some others might. It might also be due to "hardware limitation". Generally, I do not like the "filtering" (I have opposed to this in other drafts), as for me "filtering" conveys intent. If there is no intent, there is no filtering, but instead there is "dropping" or some other word.

4.7 Can we please have 4.7 that describes cases where ICMP PTB are never emitted because of misconfiguration? For instance intermediate L2 switch that has lower MTU than the L3 nodes connected to it, or mismatched MTU/MRU on two nodes connected to each other.

5.1. Can we have some kind of strong recommendation that hosts enable PLMTUD for TCP?

6. "IP encapsulations". Shouldn't this be "some packet-in-packet encapsulations"? Or does "IP encapsulations" mean "anything encapsulated in IP"? 6.3 talks about this as well, I think it's worthwile to put in a sentence that whatever is said in this document, probably applies to all kinds of encapsulations.

6.1. Err, last paragraph, aren't we getting ahead of ourselves here? I guess this is because of Geoff Hustons claims? That last paragraph is in dispute (I'd say, from talking to other people involved in DNS).

7.2. I strongly believe we need more text here. It should be something along the lines of:

"As per RFC4890, network operators MUST assure proper operation of PMTUD by making sure that PTB packets are emitted by all equipment when it can't fit a packet into a smaller MTU link, and that large MTU packets are not silently discarded due to misconfiguration. Network operators MUST NOT filter ICMP PTB packets."


As a last comment, do we know documents that tell application developers how to do what this document recommends in 5.2? If someone developers applications that use UDP for instance, how do they know what the operating system PMTUD is at any given time, to avoid the host stack fragmenting the packet? I've been interacting with people who had this specific problem, and it wasn't easy to figure out exactly how to do what is being said in this text (which I agree should be done).

Generally, I think the IETF should strongly recommend application/protocol developers to not rely on IP fragmentation, generally. So the ones listed in 6 (and I imagine there are more of them), should change the way the protocol is done. This includes DNS. So all working groups should be put on notice to start working on this problem if they don't already have a solution for it.

Mikael Abrahamsson    email: swm...@swm.pp.se

Int-area mailing list

Reply via email to