Comment 0 for bug 1806777

[impact]

in bug 1752391, cloud-init was updated to check if the initramfs has specific iscsi root device config (e.g. static initramfs config, or ibft), and if so then set up the interface in e/n/i but do not mark it as 'auto'. This prevents ifupdown from actually running dhclient for the interface; the reason for the change in that bug was to prevent a hang at shutdown, as taking down the interface removes access to the iscsi root disk (thus hanging the system).

However, iBFT has no mechanism in its spec to provide the DNS search domain, so while the interface ip address/netmask and gateway is correctly set up when the system finishes booting, it does not contain any search domain in /etc/resolv.conf. There is no way to get this information except from an actual dhcp query/response.

In the initramfs, if open-iscsi does not detect ibft (or have static config built-in) then it will not create a /run/net-IFACE.conf file; if it does find ibft (or have static config) it will create a /run/net-IFACE.conf file with the config data that it has (e.g. with ibft, all the interface configuration that ibft is able and configured to provide). After this (setting up the interface and mounting the root dev), ipconfig runs and checks for the existence of /run/net-IFACE.conf file. If it does NOT exist, then ipconfig creates it containing the result of the dhcp lease response that ipconfig received - which includes the dns search domain. However, if there is already a /run/net-IFACE.conf file, ipconfig does NOT update it. So, the file created by open-iscsi remains, without any dns search domain.

Once the system boots, resolvconf checks for existance of the /etc/net-IFACE.conf file, and if found it transfers some of the data into /etc/resolv.conf (e.g. the dns search domain). Since it does not contain any dns search domain, the /etc/resolv.conf file does not contain it either.

If the system is configured to start dhclient for the interface, then dhclient receives and sets up the system resolv.conf to use the dns search domain correctly.

For trusty: cloud-init is very different and sets up the running system ifupdown to start dhclient for the root device interface. The dns search domain is therefore configured by dhclient in the runtime system.

For xenial: cloud-init sets up the iscsi root device's interface with a e/n/i configuration for dhcp, but leaves its 'auto' line commented out, e.g.:

# control-manual ens3f0
iface ens3f0 inet dhcp
    dns-nameservers 169.254.169.254
    gateway 10.0.0.1

this results in no dns search domain for the system.

For bionic and later: cloud-init sets up netplan instead of ifupdown; netplan has no such 'auto' mechanism, but it creates systemd-networkd .network file with the "CriticalConnection" parameter set to true, which prevents systemd-networkd from ever taking the interface down, which avoids the hang on shutdown (or any other time dhcp fails/stops for the interface).

[test case]

start a xenial system that uses ibft and check /etc/resolv.conf; there will be no dns search domain.

[regression potential]

depending on the specific fix, this has the potential to cause the interface to be taken down which would then hang the system (due to root device being iscsi).

[other info]

this appears to have been introduced by bug 1752391. This bug may possibly also exist if the "ip=" parameter is used, which also triggers the cloud-init code that sets the runtime interface to 'manual' instead of 'auto', though I have not specifically tested that.