f08c810028
This runner is a little project by Bas, written in C++, that spawns
threads that then loop grabbing chunks of the (randomly shuffled but
consistently so) test list and hand it to a dEQP instance. As the
remaining list gets shorter, so do the chunks, so hopefully the
threads all complete effectively at once. It also handles restarting
after crashes automatically. I've extended the runner a bit to do
what I was doing in the bash scripts before, like the skip list and
expected failures handling. This project should also be a good
baseline for extending to handle retesting of intermittent failures.
By switching to it, we can have the swrast tests just take up one job
slot on the shared runners and keep their allotment of CPUs busy,
instead of taking up job slots with single-threaded dEQP jobs. It
will also let us (eventually, once I reprovision) switch the freedreno
runners over to threading within the job instead of running concurrent
jobs, so that memory scribbles in one pipeline don't affect unrelated
pipelines, and I can experiment with their parallelism (particularly
on a306 where we are frequently backed up) without trashing other
people's jobs.
What we lose in this process is per-test output in the log (not a big
loss, I think, since we summarize fails at the end and reducing log
length keeps chrome from choking on our logs so badly). We also drop
the renderer sanity checking, since it's not saving qpa files for us
to go poke through. Given that all the drivers involved have fail
lists, if we got the wrong renderer somehow, we'd get a job failure
anyway.
v2: Rebase on droppong of the autoscale cluster and the arm64
build/test split. Use a script to deduplicate the cts-runner
build.
v3: Rebase on the amd64 build/test container split.
Acked-by: Daniel Stone <daniels@collabora.com> (v1)
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com> (v2)
Mesa testing using gitlab-runner
The goal of the "test" stage of the .gitlab-ci.yml is to do pre-merge testing of Mesa drivers on various platforms, so that we can ensure no regressions are merged, as long as developers are merging code using the "Merge when pipeline completes" button.
This document only covers the CI from .gitlab-ci.yml and this
directory. For other CI systems, see Intel's Mesa
CI or panfrost's LAVA-based
CI (src/gallium/drivers/panfrost/ci/)
Software architecture
For freedreno and llvmpipe CI, we're using gitlab-runner on the test devices (DUTs), cached docker containers with VK-GL-CTS, and the normal shared x86_64 runners to build the Mesa drivers to be run inside of those containers on the DUTs.
The docker containers are rebuilt from the debian-install.sh script when DEBIAN_TAG is changed in .gitlab-ci.yml, and debian-test-install.sh when DEBIAN_ARM64_TAG is changed in .gitlab-ci.yml. The resulting images are around 500MB, and are expected to change approximately weekly (though an individual developer working on them may produce many more images while trying to come up with a working MR!).
gitlab-runner is a client that polls gitlab.freedesktop.org for available jobs, with no inbound networking requirements. Jobs can have tags, so we can have DUT-specific jobs that only run on runners with that tag marked in the gitlab UI.
Since dEQP takes a long time to run, we mark the job as "parallel" at some level, which spawns multiple jobs from one definition, and then deqp-runner.sh takes the corresponding fraction of the test list for that job.
To reduce dEQP runtime (or avoid tests with unreliable results), a deqp-runner.sh invocation can provide a list of tests to skip. If your driver is not yet conformant, you can pass a list of expected failures, and the job will only fail on tests that aren't listed (look at the job's log for which specific tests failed).
DUT requirements
DUTs must have a stable kernel and GPU reset.
If the system goes down during a test run, that job will eventually time out and fail (default 1 hour). However, if the kernel can't reliably reset the GPU on failure, bugs in one MR may leak into spurious failures in another MR. This would be an unacceptable impact on Mesa developers working on other drivers.
DUTs must be able to run docker
The Mesa gitlab-runner based test architecture is built around docker, so that we can cache the debian package installation and CTS build step across multiple test runs. Since the images are large and change approximately weekly, the DUTs also need to be running some script to prune stale docker images periodically in order to not run out of disk space as we rev those containers (perhaps this script).
Note that docker doesn't allow containers to be stored on NFS, and doesn't allow multiple docker daemons to interact with the same network block device, so you will probably need some sort of physical storage on your DUTs.
DUTs must be public
By including your device in .gitlab-ci.yml, you're effectively letting anyone on the internet run code on your device. docker containers may provide some limited protection, but how much you trust that and what you do to mitigate hostile access is up to you.
DUTs must expose the dri device nodes to the containers.
Obviously, to get access to the HW, we need to pass the render node
through. This is done by adding devices = ["/dev/dri"] to the
runners.docker section of /etc/gitlab-runner/config.toml.
HW 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 people can regularly use the "Merge when pipeline succeeds" button successfully (until we get marge-bot in place on freedesktop.org). As a result, we require that the test farm be able to handle a whole pipeline's worth of jobs in less than 5 minutes (to compare, the build stage is about 10 minutes, if you could get all your jobs scheduled on the shared runners in time.).
If a test farm is short the HW to provide these guarantees, consider
dropping tests to reduce runtime.
VK-GL-CTS/scripts/log/bottleneck_report.py can help you find what
tests were slow in a results.qpa file. Or, you can have a job with
no parallel field set and:
variables:
CI_NODE_INDEX: 1
CI_NODE_TOTAL: 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 consistenly 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 mesa-dev@lists.freedesktop.org after the fact explaining what happened and what the mitigation plan is for that failure next time.