[systemd-devel] high latency on systemd-resolved repsonses of LLMNR names

[Kai Krakow]

Am Mon, 10 Apr 2017 10:26:14 +0200
schrieb Lennart Poettering <lennart at poettering.net>:

> On Sun, 09.04.17 19:22, Paul Freeman (paul at coredev.org.uk) wrote:
> 
> > Hi,
> >  We are seeing high latency (>100ms) when resolving local names via
> >  LLMNR.  
> 
> Queries and responses in LLMNR are supposed to be delayed by a random
> time up to 100ms according to the RFC. See:
> 
> https://tools.ietf.org/html/rfc4795 section 2.7, and section 7.
> 
> If you add up the delay for the query and the response you'll get a
> delay up to 200ms for a full transaction.

Well it seems a bit more complicated:

The random delays are a combined value of jitter and timeout. And it
depends on whether you're currently the sender or responder: Responders
have shorter timeouts (only jitter), according to section 2.7.

It also says SHOULD wait (so it is a good idea), and it says the jitter
MAY be ignored if the responder knows the name is unique. Only the
sender MUST wait for timeout+jitter.

This is where the 200ms come from, but it may even be 1000+100ms if
you are connected to none-IEEE 802.* media, e.g. VPN or PPP interfaces
and LLMNR is configured to listen on these, as far as I understand
section 7.

According to RFC2119, the terminology SHOULD suggests that systemd could
maybe make this configurable? Maybe taking the proper warnings for this
configuration into account for administrators... Still you would see
delays of at least 100ms then because the sender MUST wait.

-- 
Regards,
Kai

Replies to list-only preferred.