Routes and IP configuration of gw0 removed when networkd restarts on a Node

[antoninbas]

Describe the bug
Thanks to @alex-vmw for reporting this issue.

When networkd crashes (or when it is restarted manually) which is something we observed on a live cluster, it seems that networkd assumes that it is managing gw0 and it therefore deletes the IP configuration for gw0, along with all the routes associated with gw0. The only solution for this at the moment is to restart the antrea-agent on the Node (or restart the Node altogether).

To Reproduce
On a Node with networkd, restart the service. Observed the gw0 configuration and the routes.

Expected
IP configuration / routes should easer be preserved or re-configured within a reasonnable time frame.

Versions:

===> Version information <===
VERSION: v0.6.0-dev
GIT_SHA: e1d56aa
GIT_TREE_STATE: clean
RELEASE_STATUS: unreleased
DOCKER_IMG_VERSION: v0.6.0-dev-e1d56aa

[antoninbas]

The original plan was to have the antrea initContainer write this file to /run/systemd/network/10-antrea.network:

[Match]
Name=gw0

[Link]
Unmanaged=yes

However, there are 2 issues with that approach:

• The host gateway may not be named gw0. In theory Antrea users can edit the config to choose a different name. This makes it hard to write the appropriate file in the initContainer.
• The entire /run directory from the host needs to be mounted in the Antrea container as we don't know if /run/systemd exists already and if it does not exist we do not want to create it.

Since this issue is specific to coreOS (see weaveworks/weave#2601), I am tempted to turn this into a "documentation" issue.

@jianjuns what do you think?

[jianjuns]

Yeah, I do not really like to get into network manager business. But I still think we should check and sync routes periodically. I think most other solutions do so.

For your issue #1 I think initContainer can mount the Agent ConfigMap and read the gateway interface there (BTW, I feel better to add an "antrea-" prefix to all interfaces we created like antrea-gw0).