[Ntp] Re: [EXT] [EXT] Focussing loop detection on loop detection only.

[Miroslav Lichvar <[e-mail]>]

On Tue, Jul 14, 2026 at 10:04:00AM +0200, David Venhoek wrote:
> The problem I see with not going after all loops is that we need to
> specify a way to decide which loops we do detect. The idea with a
> list, whilst nice in theory, has run straight into this wall in that
> we didn't have a concrete proposal how to do weighting with it last
> time, because of trouble with how to define those weights. At the same
> time, those weights are crucial for mixed-source using implementations
> like ntp-classic and ntpd-rs to work.

I didn't know there was a wall. For implementations that track and
combine time and frequency offsets independently (as chrony and
ntpd-rs do), I think the weight reported for the loop detection could
simply be the time offset weight (is that 1.0/offset_uncertainty in
ntpd-rs?). But I wouldn't want to make it a hard requirement. Maybe it
will turn out it's better to combine it with the frequency weight and
implementations should be able to do that. Or do you see some
interoperability issues coming from that? I can try to run some tests
with loops mixing different weighting, if you think it's important.

> The bloom filter, whilst not completely ideal, does deal with this
> well by just not bothering selecting which loops are important and
> which are not. I am not worried enough about malicious collisions to
> throw it out because of that, especially as most servers should (and
> will) have more carefully curated lists of sources which are trusted,
> and those sources can also cause headaches which are subtle just
> through the time sync alone, which may have more significant impact.

Yes, malicious servers can do interesting things by injecting
time/frequency offsets or faking root delay/dispersion etc. That
can be detected locally by comparing multiple sources. However, the
malicious collisions in the Bloom filter cannot be easily detected, at
least I don't see how. It can be coming from any source. Each server
would need to know the reference IDs of all servers it is directly and
indirectly serving, i.e. the same problem as loop detection itself.

-- 
Miroslav Lichvar