[PATCH] wg-quick: linux: Don't fail systemd service when using systemd-resolved

[Ronan Pigott]

On Tue Aug 27, 2019 at 4:51 PM Jason A. Donenfeld wrote:
> Could you resubmit a v2 with your Signed-off-by line?
OK.

> Instead of this hack, why not just reorder unset_dns and del_if? Are
> you worried about a race in between them if they're ordered as such?
Yeah that was my first thought, then I supposed that perhaps it was
ordered that way intentionally to prevent a situation where the dns is
unset and a client would make a dns lookup on the interface before it is
deleted. It would use the sd-resolved 'global' configured dns servers in
that case which is maybe undesired.

I realize it's a hack. I gave the explanation somewhat in the hope
someone had a better solution to the problem I was experiencing.

The real issue here is that there are there are three common
implementations of resolvconf: debian resolvconf, openresolv, and
sd-resolved compat interface. So wg-quick uses resolvconf, but is only
fully compatible with openresolv. wg-quick is (accidentally) compatible
with sd-resolved through the resolvconf compat interface, except for
this minor hiccup. This hack allows wg-quick to work for both methods
with no extra configuration from the user, no alternative code path
for sd-resolved, and no change in behavior for non sd-resolved users.

There is another way that is might be more palatable: openresolv and
sd-resolved resolvconf both support a '-f' flag for ignoring missing
interfaces, which neatly sidesteps this error. However, I don't think
debian's resolvconf supports this flag. Now that I look into it though,
it seems debian's resolvconf does not support '-l' either, which wg-quick
uses, so maybe we don't care to support this anyway. I think v2 will
just use the flag instead, if that's alright.

Users of openvpn use sd-resolved dbus interface in place of resolvconf,
such as seen here: https://github.com/jonathanio/update-systemd-resolved
but I believe this is overkill for wg-quick.