[Impact]
A widespread outage was caused on Azure instances earlier today, when systemd 237-3ubuntu10.54 was published to the bionic-security pocket. Instances could no longer resolve DNS queries, breaking networking.
For affected users, the following workarounds are available. Use whatever is most convenient.
- Reboot your instances
- or -
- Issue "udevadm trigger -cadd -yeth0 && systemctl restart systemd-networkd" as root
The trigger was found to be open-vm-tools issuing "udevadm trigger". Azure has a specific netplan setup that uses the `driver` match to set up networking. If a udevadm trigger is executed, the KV pair that contains this info is lost. Next time netplan is executed, the server loses it's DNS information.
This is the same as bug 1902960 experienced on Focal two years ago.
The root cause was found to be a bug in systemd, where if we receive a "Remove" action from a change uevent, we need to run net_setup_link(), we need to skip device rename and keep the old name.
[Testcase]
Start an instance up on Azure, any type. Simply issue udevadm trigger and reload systemd-networkd:
$ ping google.com
PING google.com (172.253.62.102) 56(84) bytes of data.
64 bytes from bc-in-f102.
$ sudo udevadm trigger && sudo systemctl restart systemd-networkd
$ ping google.com
ping: google.com: Temporary failure in name resolution
To fix a broken instance, you can run:
$ sudo udevadm trigger -cadd -yeth0 && sudo systemctl restart systemd-networkd
and then install the test packages below:
Test packages are available in the following ppa:
https:/
If you install them, the issue should no longer occur.
[Where problems could occur]
If a regression were to occur, it would affect systemd-udevd processing 'change' events from network devices, which could lead to network outages. Since this would happen when systemd-networkd is restarted on postinstall, a regression would cause widespread outages due to this SRU being targeted to the security pocket, where unattended-upgrades will automatically install from.
Side effects could include incorrect udevd device properties.
It is very important that this SRU is well tested before release.
[Other info]
This was fixed in Systemd 247 with the following commit:
commit e0e789c1e97e2cd
Author: Yu Watanabe <email address hidden>
Date: Mon, 14 Sep 2020 15:21:04 +0900
Subject: udev: re-assign ID_NET_DRIVER=, ID_NET_LINK_FILE=, ID_NET_NAME= properties on non-'add' uevent
Link: https:/
This was backported to Focal's systemd 245.4-4ubuntu3.4 in bug 1902960 two years ago. Focal required a heavy backport, which was performed by Dan Streetman. Focals backport can be found in d/p/lp1902960-
https:/
The changes between the Focal backport and the Bionic backport are:
- We use udev_device_
- device_
- We return 0 from the "if (a == DEVICE_ACTION_MOVE) " hunk instead of "goto no_rename"
- log_device_* has been changed to log_*.
See attached debdiff for Bionic backport.