This reverts a change applied in PR #18491. When interfaces are
configured by DHCP (typical in a cloud environment), ordering after
network.target cause trouble to applications expecting some network to
be present on boot (for example, cloud-init is quite brittle when
network hasn't been configured for `cloud-init.service`) and on
shutdown (for example, collectd needs to flush metrics on shutdown).
When ordering after network.target, we ensure applications relying on
network.target won't have any network reachability on boot and
potentially on shutdown.
Therefore, I think ordering before network.target is better.
We want to wait for both stacks to be active before declaring that network is active.
So either both default gateways must be specified or only IPv4 if IPv6 is disabled to
avoid dhcpcd for network-online.target.
Previously services depending on network-online.target would wait until
dhcpcd times out if it was enabled and a static network address
configuration was used. Setting the default gateway statically is enough
for the networking to be considered online.
This also adjusts the relevant networking tests to wait for
network-online.target instead of just network.target.
When dhcpcd instead of networkd is used, the network-online.target behaved
the same as network.target, resulting in broken services that need a working
network connectivity when being started.
This commit makes dhcpcd wait for a lease and makes it wanted by
network-online.target. In turn, network-online.target is now wanted by
multi-user.target, so it will be activated at every boot.
See #18319 for details. Starting network-online.target manually does not
work as it hangs indefinitely.
Additionally, don't treat avahi and dhcpcd special and sync their systemd units
with the respective upstream suggestion.
Systemd upstream provides targets for networking. This also includes a target network-online.target.
In this PR I remove / replace most occurrences since some of them were even wrong and could delay startup.
The networkd implementation sets systemd.services.dhcpcd.enable to
false in nixos/modules/tasks/network-interfaces-systemd.nix. So we need
to respect that in the dhcpcd module.
If we don't, the resumeCommand is set nevertheless, which causes the
post-resume.service to fail after resuming:
Failed to reload dhcpcd.service: Unit dhcpcd.service is masked.
post-resume.service: main process exited, code=exited, status=1/FAILURE
Failed to start Post-Resume Actions.
Dependency failed for Post-Resume Actions.
Unit post-resume.service entered failed state.
post-resume.service failed.
Signed-off-by: aszlig <aszlig@redmoonstudios.org>
This fixes several problems in the dhcpcd service:
* A segfault during startup, due to a race with udev (dhcpcd would get
an ADD event from udev, causing it to re-add an interface that it
already had, leading to a segfault later on).
* A hang/segfault processing "dhcpcd rebind" (which NixOS calls after
waking up from suspend).
Also, add "lo" to the list of ignored interfaces. It usually ignores
"lo", but apparently not when it gets an ADD event from udev.
Using pkgs.lib on the spine of module evaluation is problematic
because the pkgs argument depends on the result of module
evaluation. To prevent an infinite recursion, pkgs and some of the
modules are evaluated twice, which is inefficient. Using ‘with lib’
prevents this problem.
- Make dhcp work, use dhcpcd without udev in container
- Make login shell work, patch getty to not wait for /dev/tty0
- Make ssh work, sshd/pam do not start session
For example, the following sets up a container named ‘foo’. The
container will have a single network interface eth0, with IP address
10.231.136.2. The host will have an interface c-foo with IP address
10.231.136.1.
systemd.containers.foo =
{ privateNetwork = true;
hostAddress = "10.231.136.1";
localAddress = "10.231.136.2";
config =
{ services.openssh.enable = true; };
};
With ‘privateNetwork = true’, the container has the CAP_NET_ADMIN
capability, allowing it to do arbitrary network configuration, such as
setting up firewall rules. This is secure because it cannot touch the
interfaces of the host.
The helper program ‘run-in-netns’ is needed at the moment because ‘ip
netns exec’ doesn't quite do the right thing (it remounts /sys without
bind-mounting the original /sys/fs/cgroups).
This patch adds support for the creations of new bond devices, aggregate
pipes of physical devices for extra throughput or failover.
Additionally, add better correction at the startup of a bridge
of vlan interface (delete old, stale interfaces).
