Whoops, that last commit just wasn't very good. We need to
re-open the libvirt connection and re-get the domain object,
and also we should catch errors while destroying and undefining
the domain...
Signed-off-by: Adam Williamson <awilliam@redhat.com>
This was just an oversight - we don't destroy and undefine the
'createhdds' domain on the path where the final retry to create
an image times out, we just wipe the temp file and exit. So we
leave a qemu process sitting around until createhdds runs again
(and, on openQA production, exhausting swap space).
Signed-off-by: Adam Williamson <awilliam@redhat.com>
The way this was done, when we ran `all` on a ppc host it wiped
all the x86_64 and i686 virt-install images, and when we ran
`all` on an x86_64 host it wiped all the ppc64 virt-install
images. Doing it this way avoids that.
Signed-off-by: Adam Williamson <awilliam@redhat.com>
as today seems to be required to avoid createhdds to hang
on creation of disk_f25_minimal_2_ppc64le.img.tmp ...[13/22]
Trying to VNC connect, show anaconda menu stuck on
"Performing post-installation setup tasks"
Signed-off-by: Michel Normand <normand@linux.vnet.ibm.com>
Assumption createhdds executed on a PowerPC ppc64le host
to create the PowerPC specific images.
Detect current CPU arch of host machine to create virt-install images
only for supported architectures. (hardcoded lists)
hdds.json specific changes for PowerPC
* no desktop or kde images
Signed-off-by: Michel Normand <normand@linux.vnet.ibm.com>
'stable' means 'all current stable releases', and is used for
desktop and server as we need to ensure we have those images
available for all stable releases (including the '-3' release
while it's not EOL) for update testing. (Currently, F24 update
tests are all failing as the images are missing).
'current' means 'the current stable release' - it's the same as
'-1', but just easier to understand. It's used for support.
It seems like the installer images (in os/images) for the F26
Workstation tree somehow come from the OStree installer compose
rather than the network installer compose; install.img and
boot.iso are far larger than they should be, and match the size
of the OStree installer .iso . So instead of using those images
and bumping up the memory size to 4GiB, use the Everything tree
for F26 Workstation image builds and go back to 2GiB.
The updates.img is broken with current F26, and Radek claims the
bug is fixed. Let's see. Unfortunately I don't remember what I
put in the updates.img, so if I have to recreate it things could
get fun...
This bug is causing havoc with image creation on openqa01, so
use an updates.img for F26 image builds until it's fixed.
Signed-off-by: Adam Williamson <awilliam@redhat.com>
This is an attempt to add features desirable for creating
Taskotron base images. It extends the 'release' handling for
virt-install images in several ways to allow requesting of
'branched' and 'rawhide' image creation. It also adds an arg
to request virt-install image creation run in text mode, not
graphical mode. Graphical mode is what we always want for
openQA (so the installed OS doesn't have kernel params intended
for serial console interaction), but for Taskotron purposes,
we want the install run in text mode.
This also adds 'branched' to the default JSON file for minimal
and desktop, as we will want branched versions of these images
for the critpath update testing workflow to work on Branched
after Bodhi activation.
Signed-off-by: Adam Williamson <awilliam@redhat.com>
There's been an annoying bug on the production openQA server
for a while now: createhdds runs to refresh the upgrade base
images don't actually work most of the time, each attempt to
run virt-install fails. I finally got time to debug this a bit
today and it seems to be some kind of network issue: the VM has
no network access so the install doesn't work. I've no idea why
that's happening, but using user-mode networking seems to work
around it for now and shouldn't have any terrible side effects
even for openQA servers not affected by this problem. So let's
do this until we can pin down the real bug.
Summary:
I've come to dislike the approach where we source files that
get included in the disk images remotely; it's unnecessarily
complicated and needlessly exposes us to unexpected changes in
the remote files. So simply keep the files in this git repo
instead, none of them is very big. This also simplifies the
code.
Also, change all the kickstarts to specify --device=link and
--activate on their network lines. These arguments *should*
be included, according to the documentation (the first to
specify which device the config is for, the second to specify
that it should be activated in the installer environment), and
I think the lack of the second is actually now a problem for
the FreeIPA kickstart enrolment test (not sure exactly what
changed to make this a problem where it wasn't before, but
*something* has, and this fixes it).
Test Plan:
check that all tests still run properly. The
FreeIPA enrolment kickstart test should now get somewhat
further. Before this change, for F25 and Rawhide, it fails to
enrol at all during install. Now it should enrol properly -
that's what the kickstart changes fix, it was failing because
it wasn't using the right network config - but it still fails
when trying to log in as test1, due to RHBZ #1366403 .
Reviewers: jskladan, garretraziel
Reviewed By: garretraziel
Subscribers: tflink
Differential Revision: https://phab.qadevel.cloud.fedoraproject.org/D968
Summary:
We've kinda been having too much trouble with virt-builder
lately, mainly SELinux related issues due to how it does image
customization. It also produces images that differ in notable
ways from what a 'typical' install would give. virt-install
solves both these problems, and also gives us more flexibility
for storage configuration and post-install customization should
we need them in future.
The change isn't really too drastic, and the design is similar:
instead of virt-builder commands files, each image type now has
a kickstart file where all its customizations can be done.
There's also a single extra image dict key, 'variant', which
specifies which install tree variant to use for running the
install. It defaults to 'Everything' (for F24+) and 'Server'
(for <F24, as Everything wasn't installable until F24) but we
set it to 'Server' for the server images and 'Workstation' for
the desktop images, so those installs will use the correct
variant install class.
We run the installs in VNC. You can do it with a serial console
and log the output, but then anaconda gets clever and changes
several things in the installed system based on the fact that
you did the install over a serial console: it twiddles with
the kernel args and doesn't set graphical.target as the default.
We don't want any of that mess, so we do a VNC install.
The 'size' value is just a number of gigabytes for virt-install
images (as that's how the virt-install 'size' argument works).
This also drops some unused 32-bit images (we don't do 32-bit
KDE or Server upgrade tests, so there's no need to build those
images).
Test Plan:
Re-generate all affected images and re-run all tests
that use them, make sure they work. I am doing this on staging
at present. Note: this would render D911 unnecessary.
Reviewers: garretraziel
Reviewed By: garretraziel
Subscribers: tflink
Differential Revision: https://phab.qadevel.cloud.fedoraproject.org/D917
Summary:
Creating the virt-builder images directly with their final file
name means that if we're rebuilding them for age and the build
fails, we'll lose the existing image: it seems better to keep
it than have no image at all, when this happens. It also means
that while the rebuild is in process, the file might exist but
be useless and cause any tests that happen to be running at the
time to fail. So just like the guestfs images, create the file
with a .tmp extension initially and rename it after a successful
build.
Test Plan:
Do some image builds and check they work, check that
temp file is cleaned up on failure and ctrl-c if you can...
Reviewers: jskladan, garretraziel
Reviewed By: garretraziel
Subscribers: tflink
Differential Revision: https://phab.qadevel.cloud.fedoraproject.org/D858
Summary:
This goes along with the openqa_fedora commit to add the tests.
I didn't update the Docker instructions yet because I don't
quite remember how that goes. It might need a whole different
setup using some other networking...thing...
Test Plan: See https://phab.qadevel.cloud.fedoraproject.org/D831
Reviewers: jskladan, garretraziel
Reviewed By: garretraziel
Subscribers: tflink
Differential Revision: https://phab.qadevel.cloud.fedoraproject.org/D832
Summary:
With the previous code, the temporary file wasn't closed before
we try to upload its contents to the disk image. This seems to
cause problems when run in Python 2; the file is truncated on
upload. So rejig things a bit so we use tempfile.mkstemp, close
the file after writing to it, then remove it ourselves after
doing the upload. Tested with Python 2 and Python 3 that this
creates a correct image.
If something goes wrong at the wrong time the temp file will
be left around, but hey, it's a temp file - they get cleaned
up on system shutdown. So doesn't seem like a big deal.
Test Plan:
Check building the two images that involve a file
upload in both Python 2 and Python 3, make sure the file is
complete and correct in all cases.
Reviewers: jskladan, garretraziel
Reviewed By: garretraziel
Subscribers: tflink
Differential Revision: https://phab.qadevel.cloud.fedoraproject.org/D786
Summary:
createhdds.sh was just too damn simple and understandable, so
I thought I'd make it three times longer, object oriented,
and hard to understand!
OK, OK, that's not great sales. Alright. The main thing was to
make it smarter. This rewrite lets it do these things:
* Only create the images that are missing (not rebuild all)
* Work out the releases to build images for
* Rename images when appropriate
* Rebuild images when they need rebuilding
* Remove old / abandoned images
It can figure out what images ought to be present - including
working out the 'next' release and figuring out from that what
releases it needs images for - and build only the missing ones.
There's a 'version' concept for images; if the existing image
is older than the version given in the data file, it'll be
rebuilt. The data file can list 'rename' pairs, allowing images
to be renamed (like when we move from a single image to multiple
label/filesystem variants). This code uses fedfind's ability
to find the current release version to figure out what releases
we need virtbuilder images for (so you don't have to pass it
in). And it can find image files that aren't in the 'currently
expected' set and wipe them. Images can also have a 'maxage',
triggering a rebuild when they exceed it - this is intended
for the virtbuilder images, so we get a rebuild with the
latest updates every so often (default is two weeks).
The point of all this is to help with unattended deployment/
maintenance, i.e. the ansible deployment we have in infra;
the idea is that we can just set that up to run the 'all'
subcommand every so often, and it'll remove old images, create
new ones, and rebuild ones that are outdated.
I kept the ability to build a single image (or a whole image
'group'), and included the ability to just run a check without
actually doing a rebuild. There's a few little weird things
and holes here as it's not really the focus of the tool.
Test Plan:
Build all images, do a full test run, and see if
it works OK. Test out all the variations of building single
images / image groups, and using the 'check' command.
Reviewers: jskladan, garretraziel
Reviewed By: jskladan, garretraziel
Subscribers: tflink
Differential Revision: https://phab.qadevel.cloud.fedoraproject.org/D687
Adam Williamson
Michel Normand