GitLab provides a convenient framework for running commands in response to Git pushes.
We use it to test merge requests (MRs) before merging them (pre-merge testing),
as well as post-merge testing, for everything that hits
(this is necessary because we still allow commits to be pushed outside of MRs,
and even then the MR CI runs in the forked repository, which might have been
modified and thus is unreliable).
The CI runs a number of tests, from trivial build-testing to complex GPU rendering:
Build testing for a number of build systems, configurations and platforms
Sanity checks (
Some drivers (softpipe, llvmpipe, freedreno and panfrost) are also tested using VK-GL-CTS
Replay of application traces
A typical run takes between 20 and 30 minutes, although it can go up very quickly if the GitLab runners are overwhelmed, which happens sometimes. When it does happen, not much can be done besides waiting it out, or cancel it.
Due to limited resources, we currently do not run the CI automatically
on every push; instead, we only run it automatically once the MR has
been assigned to
Marge, our merge bot.
If you’re interested in the details, the main configuration file is
and it references a number of other files in
If the GitLab CI doesn’t seem to be running on your fork (or MRs, as they run
in the context of your fork), you should check the “Settings” of your fork.
Under “CI / CD” → “General pipelines”, make sure “Custom CI config path” is
empty (or set to the default
.gitlab-ci.yml), and that the
“Public pipelines” box is checked.
The three GitLab CI systems currently integrated are:
The Intel CI is not yet integrated into the GitLab CI. For now, special access must be manually given (file a issue in the Intel CI configuration repo if you think you or Mesa would benefit from you having access to the Intel CI). Results can be seen on mesa-ci.01.org if you are not an Intel employee, but if you are you can access a better interface on mesa-ci-results.jf.intel.com.
The Intel CI runs a much larger array of tests, on a number of generations of Intel hardware and on multiple platforms (X11, Wayland, DRM & Android), with the purpose of detecting regressions. Tests include Crucible, VK-GL-CTS, dEQP, Piglit, Skia, VkRunner, WebGL, and a few other tools. A typical run takes between 30 minutes and an hour.
CI farm expectations¶
To make sure that testing of one vendor’s drivers doesn’t block unrelated work by other vendors, we require that a given driver’s test farm produces a spurious failure no more than once a week. If every driver had CI and failed once a week, we would be seeing someone’s code getting blocked on a spurious failure daily, which is an unacceptable cost to the project.
Additionally, the test farm needs to be able to provide a short enough turnaround time that we can get our MRs through marge-bot without the pipeline backing up. As a result, we require that the test farm be able to handle a whole pipeline’s worth of jobs in less than 15 minutes (to compare, the build stage is about 10 minutes).
If a test farm is short the HW to provide these guarantees, consider dropping tests to reduce runtime. dEQP job logs print the slowest tests at the end of the run, and piglit logs the runtime of tests in the results.json.bz2 in the artifacts. Or, you can add the following to your job to only run some fraction (in this case, 1/10th) of the deqp tests.
variables: DEQP_FRACTION: 10
to just run 1/10th of the test list.
If a HW CI farm goes offline (network dies and all CI pipelines end up stalled) or its runners are consistently spuriously failing (disk full?), and the maintainer is not immediately available to fix the issue, please push through an MR disabling that farm’s jobs by adding ‘.’ to the front of the jobs names until the maintainer can bring things back up. If this happens, the farm maintainer should provide a report to email@example.com after the fact explaining what happened and what the mitigation plan is for that failure next time.
Mesa’s CI is currently run primarily on packet.net’s m1xlarge nodes (2.2Ghz Sandy Bridge), with each job getting 8 cores allocated. You can speed up your personal CI builds (and marge-bot merges) by using a faster personal machine as a runner. You can find the gitlab-runner package in Debian, or use GitLab’s own builds.
To do so, follow GitLab’s instructions to
register your personal GitLab runner in your Mesa fork. Then, tell
Mesa how many jobs it should serve (
concurrent=) and how many
cores those jobs should use (
FDO_CI_CONCURRENT=) by editing these
/etc/gitlab-runner/config.toml, for example:
concurrent = 2 [[runners]] environment = ["FDO_CI_CONCURRENT=16"]
The CI system uses Docker images extensively to cache infrequently-updated build content like the CTS. The freedesktop.org CI templates help us manage the building of the images to reduce how frequently rebuilds happen, and trim down the images (stripping out manpages, cleaning the apt cache, and other such common pitfalls of building Docker images).
When running a container job, the templates will look for an existing
build of that image in the container registry under
MESA_IMAGE_TAG. If it’s found it will be reused, and if
not, the associated .gitlab-ci/containers/<jobname>.sh` will be run
to build it. So, when developing any change to container build
scripts, you need to update the associated
a new unique string. We recommend using the current date plus some
string related to your branch (so that if you rebase on someone else’s
container update from the same day, you will get a Git conflict
instead of silently reusing their container)
When developing a given change to your Docker image, you would have to
bump the tag on each
git commit --amend to your development
branch, which can get tedious. Instead, you can navigate to the
container registry for
your repository and delete the tag to force a rebuild. When your code
is eventually merged to master, a full image rebuild will occur again
(forks inherit images from the main repo, but MRs don’t propagate
images from the fork into the main repo’s registry).
Building locally using CI docker images¶
It can be frustrating to debug build failures on an environment you don’t personally have. If you’re experiencing this with the CI builds, you can use Docker to use their build environment locally. Go to your job log, and at the top you’ll see a line like:
Pulling docker image registry.freedesktop.org/anholt/mesa/debian/android_build:2020-09-11
We’ll use a volume mount to make our current Mesa tree be what the
Docker container uses, so they’ll share everything (their build will
go in _build, according to
meson-build.sh). We’re going to be
using the image non-interactively so we use
run --rm $IMAGE
command instead of
run -it $IMAGE bash (which you may also find
useful for debug). Extract your build setup variables from
.gitlab-ci.yml and run the CI meson build script:
IMAGE=registry.freedesktop.org/anholt/mesa/debian/android_build:2020-09-11 sudo docker pull $IMAGE sudo docker run --rm -v `pwd`:/mesa -w /mesa $IMAGE env PKG_CONFIG_PATH=/usr/local/lib/aarch64-linux-android/pkgconfig/:/android-ndk-r21d/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android/pkgconfig/ GALLIUM_DRIVERS=freedreno UNWIND=disabled EXTRA_OPTION="-D android-stub=true -D llvm=disabled" DRI_LOADERS="-D glx=disabled -D gbm=disabled -D egl=enabled -D platforms=android" CROSS=aarch64-linux-android ./.gitlab-ci/meson-build.sh
All you have left over from the build is its output, and a _build directory. You can hack on mesa and iterate testing the build with:
sudo docker run --rm -v `pwd`:/mesa $IMAGE ninja -C /mesa/_build