KUnit uses Prow for presubmit and CI.
This repository contains all the code and instructions needed for generating a docker container for building and running KUnit tests on Prow, and then using that presubmit job container with a Prow cluster. Note that as of now, Prow will require a Google Cloud Storage Bucket for storing job artifacts and Gubernator for a frontend to viewing them.
The official repo for Prow is test-infra, Kubernetes' test infrastructure. It enables us to run presubmit jobs on changes to our gerrit repo against an arbitrary docker container image. We have pushed the docker image generated from here to gcr.io/kunit-presubmit/kunit. The image is pulled and deployed by our prow cluster, source is pulled into working directory of container from kunit source, and entrypoint script (kunit.sh) is ran. The script copies kunitconfig to working directory, runs kunit.py, and sends output to job artifacts that can be viewed later from the link specified in the report. Prow will comment on the Gerrit change with the status of the job (successful, a test failed, a test crashed, etc) and link to Prow URL for more details. The container is currently configured to require the inclusion of kunitconfig in repo. A convenience of this model is that updating the job image can be done without change to ProwJob configuration.
The docker image is configured in the Dockerfile. It is based off of Debian, with kernel build tools installed and the script ran by the prowjob included.
Note : The script sets kunitconfig to the latest kunitconfig at https://kunit.googlesource.com/kunitconfig at branch kunit/alpha/master.
A couple peculiarities of building/running UML in a docker container are solved in the deployment configuration discussed below under “Prow Job Specification”.
To build and push, you can either use Bazel or Docker:
Bazel allows building/pushing without dependence on docker and automates using latest kunitconfig in build. As Bazel enforces deterministic builds, all debian packages depended on are listed in debs.bzl, and based off the snapshot specified in the dpkg_src rule in WORKSPACE.
The direct dependencies needed are:
If recursive dependencies change in future snapshots, debs.bzl must be manually updated with a list of all dependencies. A simple solution:
# run interactive shell in latest debian image docker run -it --rm debian # update list of available packages apt update # generate list of all required dependencies recursively apt-cache depends --recurse --no-recommends --no-suggests --no-conflicts \ --no-breaks --no-replaces --no-enhances --no-pre-depends \ build-essential bc m4 flex bison python3 | grep "^\w" | sort -u
The snapshot used can be updated as specified in the distroless package manager bazel rules.
# build prowjob image bazel build :kunit # push to configured repo bazel run :push_kunit # build test image with kunit source and add to local docker bazel run :kunit_test # need docker to run, results to stdout docker run --privileged --tmpfs /dev/shm:exec kunit-presubmit/test:kunit_test
You must have Docker Make sure to enable sudoless docker (fixes gcr push authentication problems).
Now to build and push image :
# build image docker build . gcr.io/kunit-presubmit/kunit # confirm image is built docker images # push to gcr docker push gcr.io/kunit-presubmit/kunit
To test the container locally:
# have tmp directory with source checked out to $TMP/linux cp Dockerfile.test $TMP/Dockerfile cd $TMP # build test container which includes source docker build . -t test # run with args to handle issues with UML in Docker. docker run --privileged --tmpfs /dev/shm:exec test # extract log RUN=$(docker container ls --last 1 -q) docker cp $RUN:/artifacts/kunit.log . cat kunit.log # cleanup container docker rm $RUN
Note: Testing with an interactive shell results in unexpected behaviour. Running the UML Kernel in an entrypoint script works as intended but fails in an interactive shell.
By default, we have configured our install to use Google Container Registry. This requires installing Google Cloud SDK and configuring with project to push to. If using docker to push, you will need to enable gcloud authentication first to push with gcloud auth configure-docker
.
Prow Jobs are detailed at testinfra/prow/jobs.md. The prow job specification is held in the config.yaml. It specifies the gerrit repo for prow to poll and the specifications for the container. We found that running the container privileged and mounting a emptydir at /dev/shm fixed KUnit build errors. There may be more secure methods, but as prow doesn't expose job containers to external resources, this solution results in the cleanest Dockerfile.
We are using the decorated prow job which is recommended for all new prowjobs and detailed at test-infra/prow/pod-utilities.md. These utilities perform setup and capture output from the job container. The sidecar utility runs alongside the job container and uploads the artifacts to a GCS Bucket. Therefore, you must specify a GCS Cloud Bucket in the Plank configuration in order to upload logs / artifacts from job container. Fill in the TODO‘s in config.yaml with your bucket details. See Prow API documentation under DecorationConfig and test-infra’s default config.yaml for details.
The job script and kunitconfig stored inside in the docker image are kunit.sh and kunitconfig.
Prow sets the environment variable ARTIFACTS to specify a directory that will be exported to gcloud on job completion. Exit code of 0 signals success and 1 signals failure which will be used when reporting the Job's state to gerrit.
To deploy in any Kubernetes environment, first read Prow deployment documentation [here](https://github.com/kubernetes/test-infra/blob/master/prow/
getting_started_deploy.md) and for further clarification, the Kubernetes documentation here.
Prow comes with several components, several of which are only necessary for interacting with github webhooks. We have included here a deployment.yaml which includes just the components needed to deploy prow. Applying the configs under deployment/gerrit will deploy the gerrit adapter and the crier reporter for reporting back to gerrit. Applying the configs under deployment/lkml will deploy the lkml adapter and the custom crier reporter (named mail to allow concurrent deployment with regular crier) for handling mail.
For all Deployment objects, please check the source for the corresponding component under prow/cmd for arguments to provide the container being deployed.
For every new gerrit repo to run presubmits on, you will need to update the Gerrit component in the deployment accordingly.
Checkout prow/cluster/starter.yaml for github presubmit instructions which requires oauth token authentication. Job configuration in config.yaml does not need to change. Prow/cluster generally has default deployments for all prow components and the documentation under prow/cmd/{component} will specify the arguments to configure it.
Succinctly, a Kubernetes deployment file specifies all the API objects needed for deployment. Each prow component is a container pod that is described in an object of kind Deployment, specifying a Docker image that the deployment will launch. Services specify extra-pod communication and ServiceAccounts provide an identity to the processes that run in these pods. Roles and RoleBindings provide RBAC authorization for components. Read this article on RBAC in Kubernetes if documentation is not sufficient. PersistentVolumeClaim are a type of Volume used for maintaining state and is used for gerrit to keep track of the latest synced commit. You may also need to configure the Ingress depending on your network / deployment environment for all external communication.
If you require gerrit authentication, you will also need a git https cookie file. For a token periodically authenticated with gcloud, see/deploy grandmatriarch