Developer Documentation

Enabling new Kernel trees, builds, and tests

We can monitor different kernel trees in KernelCI. The builds and test jobs are triggered whenever the specified branches are updated. This manual describes how to enable trees in kernelci-pipeline.

Pipeline configuration

The pipeline trees configuration file has trees section. In order to enable a new tree, we need to add an entry there.

trees:
  <tree-name>:
    url: "<tree-url>"

For example,

trees:
  kernelci:
    url: "https://github.com/kernelci/linux.git"

The <tree-name> will be used in the other sections to refer to the newly added tree.

After adding a trees entry, we need to define build and test configurations for it. In the same jobs cofiguration file, the jobs section is there to specify them. ChromeOS specific job definitions are located in config/jobs-chromeos.yaml file. Depending upon the type of the job such as build or test job, different parameters are specified:

For instance,

jobs:

  <kbuild-job-name>:
    template: <job-template>
    kind: kbuild
    image: <docker-image-name>
    params:
      arch: <architecture>
      compiler: <compiler-name>
      cross_compile: <compiler-prefix>
      dtbs_check: <dtb-check-enabled>
      defconfig: <defconfig-name>
      fragments:
       - <fragment-name>
    rules:
      min_version:
        version: <kernel-version>
        patchlevel: <kernel-patch_level>
      tree:
        - <tree-name>
        - !<tree-name2>
      branch:
        - <branch-name>

  <test-job-name>:
    template: <template-name>
    kind: <kind of job, either 'job' or 'test'>
    params:
      nfsroot: <rootfs-url>
      collections: <name-of-suite>
      job_timeout: <timeout>
    kcidb_test_suite: <kcidb-mapping>
    rules:
      min_version:
        version: <kernel-version>
        patchlevel: <kernel-patch-level>
      tree:
        - <tree-name>
        - !<tree-name2>
        - <tree-name3>:<branch-name>

Here is the description of each field:

• template: A jinja2 template should be added to the config/runtime directory. This template will be used to generate the test definition.
• kind: The kind field specifies the type of job. It should be kbuild for build jobs, job for a test suite, and test for a single test case.
• image: The image field specifies the Docker image used for building and running the test. This field is optional. For example, LAVA test jobs use an image defined in the test definition template instead.
• params: The params field includes parameters for building the kernel (for kbuild jobs) or running the test. These parameters can include architecture, compiler, defconfig options, job timeout, etc.
• rules: The rules field defines job rules. If a test should be scheduled for a specific kernel tree, branch, or version, these rules can be specified here. The rules prefixed with ! exclude the specified condition from job scheduling. For example, in the given scenario, the scheduler does not schedule a job if an event is received for the kernel tree tree-name2. It is also possible to schedule job for a specific branch of a tree. For example, a rule tree:branch can be mentioned under tree or branch section to define that, still accepting the ! prefix to ensure jobs can be scheduled on all branches from a given tree, except a few select ones. Rules can be defined in relation to any attribute present in the parent node (such as defconfig or fragments or even arch for test jobs having a parent node of kbuild type). min_version and max_version are special cases to allow jobs to run only on given kernel versions.
• kcidb_test_suite: The kcidb_test_suite field maps the KernelCI test suite name with the KCIDB test. This field is not required for build jobs (kind: kbuild). When adding new tests, ensure their definition is present in the tests.yaml file in KCIDB.

Common patterns are often defined using YAML anchors and aliases. This approach allows for concise job definitions by reusing existing configurations. For example, a kbuild job can be defined as follows:

  kbuild-gcc-12-arm64-preempt_rt_chromebook:
    <<: *kbuild-gcc-12-arm64-job
    params:
      <<: *kbuild-gcc-12-arm64-params
      fragments:
       - 'preempt_rt'
       - 'arm64-chromebook'
      defconfig: defconfig
    rules:
      tree:
      - 'stable-rt'

The test job example is:

  kselftest-exec:
    template: generic.jinja2
    kind: job
    params:
      test_method: kselftest
      nfsroot: 'http://storage.kernelci.org/images/rootfs/debian/bookworm-kselftest/20240313.0/{debarch}'
      collections: exec
      job_timeout: 10
    kcidb_test_suite: kselftest.exec

Please have a look at config/jobs.yaml and config/jobs-chromeos.yaml files to check currently added job definitions for reference.

We need to specify which branch to monitor of a particular tree for trigering jobs in build_configs.

build_configs:
  <name-of-variant0>:
    tree: <tree-name>
    branch: <branch-name0>

  <name-of-variant1>:
    tree: <tree-name>
    branch: <branch-name1>

That’s it! The tree is enabled now. All the jobs defined under jobs section of config file would run on the specified branched for this tree.

Schedule the job

We also need a scheduler entry in config/scheduler.yaml for the newly added job to specify pre-conditions for scheduling, and defining runtime and platforms for job submissions.

For example,

scheduler:

  - job: <kbuild-job-name>
    event: <API-pubsub-event>
    runtime:
      name: <runtime-name>

  - job: <test-job-name>
    event: <API-pubsub-event>
    runtime:
      type: <runtime-type>
      name: <runtime-name>
    platforms:
      - <device-type>

Here is the description of each field:

• job: Specifies the job name, which must match the name used in the jobs section.
• event: Specifies the API PubSub event triggering the test scheduling. For example, to trigger the kbuild job when new source code is published, the event is specified as:

  event:
    channel: node
    name: checkout
    state: available

For a test that requires a successful completion of a build job such as kbuild-gcc-10-arm64, specify the event as follows:

  event:
    channel: node
    name: kbuild-gcc-10-arm64
    state: available

For a post-processing job that requires completion of all test jobs for a given build job such as kbuild-gcc-10-arm64, specify the event as follows:

  event:
    channel: node
    name: kbuild-gcc-10-arm64
    result: pass
    state: done

Here, node refers to the name of API PubSub channel where node events are published.

• runtime: Select a runtime for scheduling and running the job. Supported runtimes include shell, docker, lava, and kubernetes. Specify the runtime type from the runtimes section. Note that the name property is required for lava and kubernetes runtimes to specify which lab or Kubernetes context should execute the test. Several LAVA labs (such as BayLibre, Collabora, Qualcomm) and Kubernetes contexts have been enabled in KernelCI.
• platforms: Includes a list of device types on which the test should run. These should match entries defined in the platforms section, such as qemu-x86, bcm2711-rpi-4-b, and others.

After following these steps, run your pipeline instance to activate your newly added test configuration.

An example of enabling a new job

Here is an example of enabling kselftest (found in <linux_kernel>/tools/testing/selftests) job. Most parameters are already set for other kselftests. Those must be reused while enabling new kselftests which makes enabling new tests pretty straight forward. The dt kselftest suite can be enabled by:

1. Create a job definition in jobs section.

  kselftest-dt:
    template: generic.jinja2
    kind: job
    params:
      test_method: kselftest
      nfsroot: 'http://storage.kernelci.org/images/rootfs/debian/bookworm-kselftest/20240221.0/{debarch}'
      collections: dt
      job_timeout: 10
    rules:
      min_version:
        version: 6
        patchlevel: 7
    kcidb_test_suite: kselftest.dt

2. Create a scheduler entry for kselftest-dt job. Let’s schedule this suite in lava-collabora lab at kbuild-gcc-12-arm64 build completion event. The entry by using already defined anchors would be:

  - job: kselftest-dt
    event: *kbuild-gcc-12-arm64-node-event
    runtime: *runtime-lava-collabora
    platforms:
      - bcm2711-rpi-4-b

There can be multiple entries for a suite in scheduler section with different events, runtime or platform.

Test hierarchy

Once you enable tests for your desired Kernel tree, maestro will start running tests and storing results in maestro DB. This section explains how maestro structures test results hierarchy.

Job vs Test

We can have a test suite or a single test case running in maestro. Conceptually, we use the term job for test suites and test for a single test case.

For instance, Kernel version check i.e. kver test is a single test returning directly pass or fail result. Hence, kver would directly represent the result and won’t have any child tests running. That’s why we use node kind test for such tests.

Whereas we also have different test suites running. For example, kunit, kselftest, rt-tests, baseline (boot tests). These suites will have child tests or child jobs. All the nodes with child jobs or test cases would be represented with node kind job.

Please check the configuration file for various job and test definitions. Jobs will have kind: job and test cases will have kind: test respectively.

Visualize hierarchy

Here is a visual representation of test hierarchy in maestro. As mentioned earlier, nodes with child jobs or tests will have kind=job and test cases without any child node will have kind=test.

For example, kunit test hierarchy would look like below:

graph TD
    A["kunit <br> (kind: job)"] --> B("config <br> (kind: test)")
    A --> C["build <br> (kind: test)"]
    A -->  D["exec <br> (kind: job)"]
    D --> E["input_core <br> (kind: job)"]
    E --> F["input_test_polling <br> (kind: test)"]

In the above example, the job node kunit is created first. It has a child job and a test node. exec is a job node as it contains child test cases, whereas build and config are test cases without any sub-tests.

LAVA job hierarchy

As stated earlier, maestro supports different runtimes for tests, such as docker, kubernetes, and LAVA.

We treat LAVA jobs a bit differently with respect to other jobs. All the LAVA job nodes will have a setup test suite that runs pre-condition checks (e.g. test device login or generate tarball for tast test) before running the actual test definition.

Here is the test hierarchy for kselftest-cpufreq job:

graph TD
    A["kselftest-cpufreq <br> (kind: job)"] --> B("setup <br> (kind: job)")
    A --> C("kselftest-cpufreq <br> (kind: job)")
    C --> D["cpufreq_main_sh <br> (kind: test)"]
    C --> E["shardfile-cpufreq <br> (kind: test)"]
    B --> F["login <br> (kind: test)"]
    B --> G["kernelmsg <br> (kind: test)"]

In the above example, root node kselftest-cpufreq is the main job node. It has two child nodes. One is a setup test suite as mentioned above, and another is a job node with the same name i.e. kselftest-cpufreq which is the actual test definition. kselftest-cpufreq (child job node) will be executed only after setup test suite’s successful execution.

Checkout result evaluation

Maestro pipeline creates a checkout node i.e., root node whenever a new kernel revision is detected. All the build and test jobs are created as child nodes of this root node.

Below is the explanation of checkout node result evaluation based on different conditions.

Tarball service creates/updates a local git repo and uses it to create a kernel source tarball for the new revision. It transits checkout node to result=fail if git repo update operation fails.

flowchart TD
  subgraph tarball_service[Tarball Service]
    update_repo{Update git <br /> repo failed?}
    update_repo --> |Yes| fail_checkout[Set 'checkout' node <br />state=done, result=fail]
  end

Timeout service is responsible for detecting timed-out nodes and transiting them to appropriate state/result based on child job status. Below is the visual representation of how the checkout node is handled by the service based on its state:

flowchart TD
  subgraph timeout_service[Timeout Service]
    node_timedout{Checkout node <br /> timed out?}
    node_timedout --> |state=running| running_checkout[Set 'checkout' node <br />state=done, result=incomplete]
    node_timedout --> |state=available/closing| available_checkout[Set 'checkout' node <br />state=done, result=pass]
  end

Test result evaluation

This section is intended to explain how results are evaluated in maestro. For test cases (job nodes with kind: test), it directly uses test results i.e. pass, fail, or skip from the data received from the respective job.

In case of a test suite (job node with kind: job), maestro evaluates result based on child test suites and tests.

Let’s take an example of a job_node with the following hierarchy:

graph TD
    A[job_node] --> B(test_case_1)
    A --> C(test_case_2)
    A --> D(test_suite_1)
    D --> E(test_case_3)
    D --> F(test_case_4)

Here is the visual representation of the result evaluation of the job_node:

graph TD
    A(test_case_1)
    B(test_case_2)
    C(test_case_3)
    D(test_case_4)
    C --> E{All tests passed?}
    D --> E
    E --> |Yes|F('test_suite_1' result <br> set to 'pass')
    E --> |No|G{One of the tests failed?}
    G --> |Yes|H('test_suite_1' result <br> set to 'fail')
    G --> |No|I{All tests skipped?}
    I --> |Yes|J('test_suite_1' result <br> set to 'skip')
    A --> K{All tests and test suites passed?}
    B --> K
    F --> K
    H --> K
    J --> K
    K --> |Yes|L('job_node' result <br> set to 'pass')
    K --> |No|M{One of immediate <br> child suite or test failed?}
    M --> |Yes|N{'setup' test suite exists <br> and failed?}
    M --> |No|Q{All immediate <br> child suites or tests skipped?}
    N --> |Yes|O('job_node' result <br> set to 'incomplete')
    N --> |No|P('job_node' result <br> set to 'fail')
    Q --> |Yes|R('job_node' result <br> set to 'skip')

The summary of how the result evaluated is: A job node result will be set to

• pass if all its tests/test suites passed
• fail if one of its tests/test suites failed
• skip if all tests/test suites are skipped
• incomplete if setup test suite failed