[dns-operations] Verisign TLDs, some other servers may trim critical glue from very large referrals

[Florian Weimer]

* Stephane Bortzmeyer:

> On Mon, Jan 07, 2019 at 07:59:32AM +0100,
>  Daniel Stirnimann <daniel.stirnimann at switch.ch> wrote 
>  a message of 21 lines which said:
>
>> I think they want to protect from DNS spoofing attacks with fragments
>> [1] and they believe setting the EDNS0 buffer size to 512 bytes reduces
>> the attack vector:
>
> No article (not the Shulman paper, nor the articles on the Let's
> Encrypt site) discuss 512 vs 1280, just fragmentation
> vs. non-fragmentation. 1280 will never fragment with IPv6 and very
> rarely with IPv4. Therefore, I still don't understand 512.

Some authoritative servers honor ICMP requests to lower the path MTU to
very small values (which is why I think a client-side workaround is
rather incomplete).  512 is just the lowest value you can use.

> Also, the proper protection against the Shulman fragmentation attack
> is DNSSEC.

This is not something a CA can enable, though.

Thanks,
Florian