When using iproute2's ip binary, you can omit the dev parameter, e.g. ip link set up eth0 instead of ip link set up dev eth0.
This breaks if for some reason your device is named e.g. he, hel, … because it is interpreted as ip link set up help.
I just encountered this bug using networking.bridges trying to create an interface named he.
I used a grep on nixpkgs to try to find iproute2 invocations using variables without the dev keyword, and found a few, and fixed them by providing the dev keyword.
I merely fixed what I found, but the use of abbreviated commands makes it a bit hard to be sure everything has been found (e.g. ip l set … up instead of ip link set … up).
The networkd backend logic for setting DHCP= on an interface is bugged
and inconsistent with the scripted logic. Consider this simple NixOS
configuration:
{
networking.useNetworkd = true;
networking.interfaces.eth0.wakeOnLan.enable = true;
}
The default value of networking.useDHCP is true, so we expect our eth0
interface to have DHCP enabled. With the scripted backend, this works.
But the networkd backend generates the following 40-eth0.network file:
[Match]
Name=eth0
[Network]
DHCP=no
IPv6PrivacyExtensions=kernel
This is happening because the wakeOnLan configuration creates a key in
networking.interfaces, and the networkd backend erroneously checks that
instead of for explicitly configured IP addresses as in the scripted
backend. The documentation is also inconsistent across various options.
This change aligns the networkd backend and option documentation to the
actual behavior of the scripted backend, and updates a test to account
for this behavior for both backends.
The logic for configuring a gateway without an interface specified adds
a route with Gateway= to *every interface* configured by NixOS for
networkd. This leads to nonsensical configurations like the following:
[Network]
DHCP=no
Address=192.168.0.1/24
[Route]
Gateway=10.0.0.1
GatewayOnLink=false
We remove this logic and make defaultGateway.interface required to
configure a default gateway when using networkd.
We can ignore the removal of GatewayOnLink because systemd defaults it
to "no" anyway.
When interface and address are both specified, we can set Gateway= on
the named interface. The existing logic assumes interface is not set
(since it's guarded by assertion) so we now disable it when interface
has a value.
As a bonus, we now support the defaultGateway.metric option when
interface is set.
A further bug to our strange multi-user.target depending on
network-online.target issue is that systemd recently changed the
behaviour of systemd-networkd-wait-online to no longer consider the
absence of interfaces with RequiredForOnline to be sufficient to be
online: https://github.com/systemd/systemd/pull/27825
On the advice of the systemd developers
(https://github.com/systemd/systemd/issues/29388), this commit changes
the configuration of systemd-networkd-wait-online to pass --any by
default, and lets the default DHCP interfaces be RequiredForOnline
as they would be by default if the option is omitted.
It is plausible that systemd-networkd-wait-online may still fail if
there are no interfaces at all. However, that probably cannot be
avoided.
systemd.network(5) describes Domains= as a "list of domains which should
be resolved using the DNS servers on this link." This setting is read by
systemd-resolved.service, and it's used to configure both search domains
and DNS query routing.
Adding the search domains from `networking.search` is unnecessary
because these are already configured globally in `resolved.conf` through
the default value of `services.resolved.domains`.
Adding the system's `networking.domain` to each network is unexpected
and probably incorrect. A user may not expect that the domain is in
effect automatically added to the search domains even if not specified
in `networking.search`.
Both of these network-level assignments are problematic in cases where
the NixOS networkd module is not managing every interface on the system.
In that scenario, the managed interfaces will have Domains= set while
the others do not. That will cause systemd-resolved to route DNS queries
for the search domains and the system domain to only those managed
interfaces.
While it might seem odd, 0.0.0.0/0 or ::/0 gateways are valid and
commonly used on point-to-point links (e.g. a wireguard tunnel) to
indicate that all traffic needs to be sent to a given interface.
systemd-networkd actually documents this as a valid configuration in its
man pages [1].
Tested to do the right thing in one of my NixOS containers using
a Wireguard tunnel as its default route.
[1] https://www.freedesktop.org/software/systemd/man/systemd.network.html#DefaultRouteOnDevice=
Previously this wasn't done in the `forEach`-expression for
`cfg.interfaces` and thus `networking.useDHCP` didn't have any effect if
no further interface was statically configured.
Commit ca58bd0a50 broke the test networking.networkd.static. This happened because the test sets `networking.defaultGateway`. This is implemented by adding the gateway to the list of `routes` using `mkDefault`. The `routes` are then overridden by an empty list in the newly added code. Replace `mkDefault` with `id` so the two lists are merged and everything (hopefully) works as expected.
See https://github.com/NixOS/nixpkgs/pull/144590 for a more context.
allows configuration of foo-over-udp decapsulation endpoints. sadly networkd
seems to lack the features necessary to support local and peer address
configuration, so those are only supported when using scripted configuration.
This reverts commit d349582c07.
The workaround initially applied isn't necessary anymore, as 247.3
contains the following commit:
> 242fc1d261 network: fix IPv6PrivacyExtensions=kernel handling
… which fixes https://github.com/systemd/systemd/issues/18003.
Systemd dependencies for scripted mode
were refactored according to analysis in #34586.
networking.vswitches can now be used with systemd-networkd,
although they are not supported by the daemon, a nixos receipe
creates the switch and attached required interfaces (just like
the scripted version).
Vlans and internal interfaces are implemented following the
template format i.e. each interface is
described using an attributeSet (vlan and type at the moment).
If vlan is present, then interface is added to the vswitch with
given tag (access mode). Type internal enabled vswitch to create
interfaces (see openvswitch docs).
Added configuration for configuring supported openFlow version on
the vswitch
This commit is a split from the original PR #35127.
Systemd dependencies for scripted mode
were refactored according to analysis in #34586.
networking.vswitches can now be used with systemd-networkd,
although they are not supported by the daemon, a nixos receipe
creates the switch and attached required interfaces (just like
the scripted version).
Vlans and internal interfaces are implemented following the
template format i.e. each interface is
described using an attributeSet (vlan and type at the moment).
If vlan is present, then interface is added to the vswitch with
given tag (access mode). Type internal enabled vswitch to create
interfaces (see openvswitch docs).
Added configuration for configuring supported openFlow version on
the vswitch
This commit is a split from the original PR #35127.
Just maching all network interfaces caused many breakages, see #18962
and #71106.
We already don't support the global networking.useDHCP,
networking.defaultGateway(6) options if networking.useNetworkd is
enabled, but direct users to configure the per-device
networking.interfaces.<name?>.… options.
With systemd version 243 network units with empty match block will
generate warnigs. The reasoning seems to be that the intended behaviour
is hard to infere. Being explicit about really meaning any interface is
the reasonable thing here.
We want to get rid of this mechanism in the long run but as long as we
do not have a replacement we should stick with it and keep it in
reasonable good shape.
With systemd v242 using the `Gateway` attribute of the `[Network]`
section will lead to "onlink" routes on all the device that are matched
by the default configuration (typically all devices) causing multiple
default routes (even on localhost).
We can only avoid that - while keeping our default route option - when
we mark the route as explicitly not on link. Only gateways that are
within a subnet of one of the assigned interface addresses will be
installed into the routing table.
