Re: Getting "write: address family not supported by protocol family" when using Restic over Tailscale

[Greg Troxel <gdt%lexort.com@localhost>]

Matti Andreas Kähäri <andreas.kahari%abc.se@localhost> writes:

> I have two "sibling" VM's on the same host ("henry", Alpine Linux), one
> is "eeyore" (OpenBSD) and another is "mzulft" (NetBSD). "Sibling" means
> I can test over the same network distance as they are next to each
> other.
>
> On NetBSD:
> mzulft% iperf3 -c henry -t 5
> [ ID] Interval           Transfer     Bitrate       
> [  7]   0.00-5.02   sec  60.8 MBytes   101 Mbits/sec
> [  7]   0.00-5.03   sec  60.8 MBytes   101 Mbits/sec

> On OpenBSD:
> eeyore% iperf3 -c henry -t 5
> [ ID] Interval           Transfer     Bitrate         Retr
> [  4]   0.00-5.02   sec   349 MBytes   583 Mbits/sec    0
> [  4]   0.00-5.02   sec   349 MBytes   583 Mbits/sec

That is different and there are many possible reasons. I don't see that
this implicates tailscale, or anything in particular.  At least it's
well into usable rather than too painful to put up with.

> The difference is not that big, but NetBSD still clocks in at 100+ Mbit/s
> slower than macOS.

But you do not have bare metal NetBSD on a comparable-cpu machine with a
decent network card.

> All NetBSD systems tested has the fix to if_tun.c applied (thanks for
> that!).  But, as I said, I'm not personally greatly troubled by this
> slight slowness.  I'm just happy things work.

Understood about it not being a big deal.

In the above tests, I don't see how tailscale/tun is used, as you are on
a vm guest to vm, or on the same lan.  But it's ok that I don't
understand.

To investigate, two approaches, neither easy:
  - use tcpdump to record a single test tcp transfer, and then visualize
    the details to see if there is any pathological
  - understand the interface types, and research issues with virtual and
    physical interfaces