It turns out that this breaks ipv6 config with NM. Instead what we want
is for glean to not up interfaces on boot (see the depends-on).
Change-Id: I6c5bc76c433e29f02d3266ab8f669015125ec954
Depends-On: https://review.opendev.org/#/c/688031
By default network-scripts package isn't installed, so the directories
for these files don't exist either. Skip by default for Centos 8.
Change-Id: I194ec3735e17f27e586386541dc51f775b01e510
Use the wrapper calls from Ia267a60eecfa8f4071dd477d86daebe07e9a7e38
to install glean.
Using this wrapper means we cover all cases without more and more
branches; it should work for python2, python3 and also the special
case of RHEL/CentOS where dib-python points to the special
/usr/libexec/platform-python (which is python3.6 with inbuilt pip)
Change-Id: If624e8bb66ce0761fc0d5f34c2bed8b93a7daeee
NetworkManager with simple-init has proven to be stable in OpenStack
infra, switch to it by default for CentOS and Fedora. For CentOS 8
and Fedora, add a check to make it the only option. Thus only CenOS 7
remains optionally using the legacy scripts; this is likely not used
anywhere (infra is really the primary user, where NetworkManager is
already used); we can likely remove this variable (and hence path) in
a future cleanup.
In the setup, remove rhel7 element which was never really tested.
Reorganise the fallthrough to call out the default paths as doing
nothing.
Change-Id: Ic996956da4b85f7d95179b8df9881d5f52c091af
This is a follow-on to I475a253091cbaf63687b91c748c31a6753bb0f57 as we
are still seeing issues on some clouds with unconfigured networking.
We increase the timeout, but also make it configurable so we can
fiddle it without a dib release in the gate.
To follow-on from the experimentation done by clarkb, I can confirm by
emperical testing on a Centos 7 image (from today, today being this
change's date) that setting
net.ipv6.conf.all.autoconf=0
by itself is "fatal" and the interfaces do not come up; i.e. nm does
not by default seem to re-enable ipv6 for the interface. However,
explicitly adding:
IPV6INIT=yes
IPV6_AUTOCONF=yes
to the interface file *does* seem to make it work, even if
"all.autoconf=0" is set (then again, there's also bugs about the
effect of this [1]). However, no extant distribution (I can currently
find) does anything like this by default.
If this continues, this may be an option. Another might be to avoid
the use of the nm-settings-ifcfg-rh profiles and move directly to nm
ini files with glean.
[1] https://bugzilla.kernel.org/show_bug.cgi?id=11655
Change-Id: I869ebffc8cde3bbff573f6583fd9dd02a5598590
The linux kernel and NetworkManager fight each other over control for
interface management when router advertisements are in use. Long story
short if the linux kernel configures a network interface for ipv6
before NetworkManager attempts to manage that interface then NM will
ignore the interface and not configure ipv4 on it.
This can happen because the kernel is configured to send router
advertisements solicitations which result in router advertisements which
the kernel uses to configure the interface(s). There is a default of a 1
second delay before sending the solicitation which in many cases is long
enough that NM has started before then. However, in slower environments
like those used for testing with qemu this isn't long enough.
Some testing by hand indicates that 15 seconds is about right so
increase the delay to 15 seconds via sysctl.conf.
Note this may increase boot times in ipv6 only environments (though it
is hard to be sure due to how systemd starts everything at once and does
socket activation and the like).
Change-Id: I475a253091cbaf63687b91c748c31a6753bb0f57
There are several jobs depending on working opensuse 15.1
images in nodepool, so it makes sense to ensure its working.
Also upgrade the previously marked experimental opensuse-15.0
job as it tests the xenial->opensuse combination, which is
particularly difficult to keep working and we'll need it in
the CI.
Change-Id: Icb6d998756ce5221e017959dcb59b21f0f023454
Depending on the version of $DIB_PYTHON_VERSION, we can either use pip /
pip3 to install glean. This is helpful for newer OSes that might not
want to ship python2 (pip).
Change-Id: I25c5927a1eb55ee16b919dd64403184f335839b6
Signed-off-by: Paul Belanger <pabelanger@redhat.com>
This plumbs through an "--use-nm" flag to glean which instructs it to
setup interface bringup with NetworkManager rather than legacy network
enablement scripts.
In this case, install the NetworkManager package. In the non-nm case,
also install the network-scripts for Fedora 29 -- this has stopped
being installed by default (it's been deprecated since forever).
As noted in the docs, this is currently really only relevant on the
supported rpm distros which are using the ifcfg-rh NetworkManager
plugin to effectively re-use old config files. However,
NetworkManager has similar plugins for other platforms, so support can
be expanded if changes are proposed.
Depends-On: https://review.openstack.org/618964
Change-Id: I4d76e88ce25e5675fd5ef48924acd09915a62a4b
According to http://bit.ly/2HA4oDO and
the official Ubuntu manual
http://manpages.ubuntu.com/manpages/xenial/man5/interfaces.5.html
source-dir support has been removed from Ubuntu >= 16.04/Xenial
Once an image is generated and booted, moving the dhcp interface(s)
declaration(s) from /etc/network/interfaces into specific subentries
of /etc/network/interfaces.d and calling 'service networking restart'
just make your instance unreachable and all interfaces are left
unconfigured.
This patchset fixes this issue
Change-Id: I6b6b99c81490c874c5db5405c2fbf3c180c87464
When a glean is running on centos with multiple NICs, it will try to
systemctl enable network.service multiple times for each interface.
Because of systemd magic, it is possible for the systemctl command to
fail in a race condition.
glean shouldn't be enabling network.service during boot in
pre-networking phases (Ib2b618dd975ca44e9c6b0a2c9027642ffc46b9b0). I
have proposed I8319f1ed6498a9d447950c2b4b34bca59e7b97e4 to remove this
and document the behaviour.
This also bring across suse's version
(I20bffabd333ea290d8712ec2a467f2b2d5678f3a)
Change-Id: I89d9443cb61e287bd0d9da3f48315272218ee335
Signed-off-by: Paul Belanger <pabelanger@redhat.com>
Currently we have all our elements and library files in a top-level
directory and install them into
<root>/share/diskimage-builder/[elements|lib] (where root is either /
or the root of a virtualenv).
The problem with this is that editable/development installs (pip -e)
do *not* install data_files. Thus we have no canonical location to
look for elements -- leading to the various odd things we do such as a
whole bunch of guessing at the top of disk-image-create and having a
special test-loader in tests/test_elements.py so we can run python
unit tests on those elements that have it.
data_files is really the wrong thing to use for what are essentially
assets of the program. data_files install works well for things like
config-files, init.d files or dropping documentation files.
By moving the elements under the diskimage_builder package, we always
know where they are relative to where we import from. In fact,
pkg_resources has an api for this which we wrap in the new
diskimage_builder/paths.py helper [1].
We use this helper to find the correct path in the couple of places we
need to find the base-elements dir, and for the paths to import the
library shell functions.
Elements such as svc-map and pkg-map include python unit-tests, which
we do not need tests/test_elements.py to special-case load any more.
They just get found automatically by the normal subunit loader.
I have a follow-on change (I69ca3d26fede0506a6353c077c69f735c8d84d28)
to move disk-image-create to a regular python entry-point.
Unfortunately, this has to move to work with setuptools. You'd think
a symlink under diskimage_builder/[elements|lib] would work, but it
doesn't.
[1] this API handles stuff like getting files out of .zip archive
modules, which we don't do. Essentially for us it's returning
__file__.
Change-Id: I5e3e3c97f385b1a4ff2031a161a55b231895df5b
