3 Understanding the Yocto Project Autobuilder

3.1 Execution Flow within the Autobuilder

The “a-full” and “a-quick” targets are the usual entry points into the Autobuilder and it makes sense to follow the process through the system starting there. This is best visualized from the Autobuilder Console view (https://autobuilder.yoctoproject.org/typhoon/#/console).

Each item along the top of that view represents some “target build” and these targets are all run in parallel. The ‘full’ build will trigger the majority of them, the “quick” build will trigger some subset of them. The Autobuilder effectively runs whichever configuration is defined for each of those targets on a separate buildbot worker. To understand the configuration, you need to look at the entry on config.json file within the yocto-autobuilder-helper repository. The targets are defined in the ‘overrides’ section, a quick example could be qemux86-64 which looks like:

"qemux86-64" : {
      "MACHINE" : "qemux86-64",
      "TEMPLATE" : "arch-qemu",
      "step1" : {
            "extravars" : [
                  "IMAGE_FSTYPES:append = ' wic wic.bmap'"
                 ]
     }
},

And to expand that, you need the “arch-qemu” entry from the “templates” section, which looks like:

"arch-qemu" : {
      "BUILDINFO" : true,
      "BUILDHISTORY" : true,
      "step1" : {
            "BBTARGETS" : "core-image-sato core-image-sato-dev core-image-sato-sdk core-image-minimal core-image-minimal-dev core-image-sato:do_populate_sdk",
      "SANITYTARGETS" : "core-image-minimal:do_testimage core-image-sato:do_testimage core-image-sato-sdk:do_testimage core-image-sato:do_testsdk"
      },
      "step2" : {
            "SDKMACHINE" : "x86_64",
            "BBTARGETS" : "core-image-sato:do_populate_sdk core-image-minimal:do_populate_sdk_ext core-image-sato:do_populate_sdk_ext",
            "SANITYTARGETS" : "core-image-sato:do_testsdk core-image-minimal:do_testsdkext core-image-sato:do_testsdkext"
      },
      "step3" : {
            "BUILDHISTORY" : false,
            "EXTRACMDS" : ["${SCRIPTSDIR}/checkvnc; DISPLAY=:1 oe-selftest ${HELPERSTMACHTARGS} -j 15"],
            "ADDLAYER" : ["${BUILDDIR}/../meta-selftest"]
      }
},

Combining these two entries you can see that “qemux86-64” is a three step build where the bitbake BBTARGETS would be run, then bitbake SANITYTARGETS for each step; all for MACHINE="qemx86-64" but with differing SDKMACHINE settings. In step 1 an extra variable is added to the auto.conf file to enable wic image generation.

While not every detail of this is covered here, you can see how the template mechanism allows quite complex configurations to be built up yet allows duplication and repetition to be kept to a minimum.

The different build targets are designed to allow for parallelization, so different machines are usually built in parallel, operations using the same machine and metadata are built sequentially, with the aim of trying to optimize build efficiency as much as possible.

The config.json file is processed by the scripts in the Helper repository in the scripts directory. The following section details how this works.

3.2 Autobuilder Target Execution Overview

For each given target in a build, the Autobuilder executes several steps. These are configured in yocto-autobuilder2/builders.py and roughly consist of:

  1. Run clobberdir.

    This cleans out any previous build. Old builds are left around to allow easier debugging of failed builds. For additional information, see clobberdir.

  2. Obtain yocto-autobuilder-helper

    This step clones the yocto-autobuilder-helper git repository. This is necessary to prevent the requirement to maintain all the release or project-specific code within Buildbot. The branch chosen matches the release being built so we can support older releases and still make changes in newer ones.

  3. Write layerinfo.json

    This transfers data in the Buildbot UI when the build was configured to the Helper.

  4. Call scripts/shared-repo-unpack

    This is a call into the Helper scripts to set up a checkout of all the pieces this build might need. It might clone the BitBake repository and the OpenEmbedded-Core repository. It may clone the Poky repository, as well as additional layers. It will use the data from the layerinfo.json file to help understand the configuration. It will also use a local cache of repositories to speed up the clone checkouts. For additional information, see Autobuilder Clone Cache.

    This step has two possible modes of operation. If the build is part of a parent build, it’s possible that all the repositories needed may already be available, ready in a pre-prepared directory. An “a-quick” or “a-full” build would prepare this before starting the other sub-target builds. This is done for two reasons:

    • the upstream may change during a build, for example, from a forced push and this ensures we have matching content for the whole build

    • if 15 Workers all tried to pull the same data from the same repos, we can hit resource limits on upstream servers as they can think they are under some kind of network attack

    This pre-prepared directory is shared among the Workers over NFS. If the build is an individual build and there is no “shared” directory available, it would clone from the cache and the upstreams as necessary. This is considered the fallback mode.

  5. Call scripts/run-config

    This is another call into the Helper scripts where it’s expected that the main functionality of this target will be executed.

3.3 Autobuilder Technology

The Autobuilder has Yocto Project-specific functionality to allow builds to operate with increased efficiency and speed.

3.3.1 clobberdir

When deleting files, the Autobuilder uses clobberdir, which is a special script that moves files to a special location, rather than deleting them. Files in this location are deleted by an rm command, which is run under ionice -c 3. For example, the deletion only happens when there is idle IO capacity on the Worker. The Autobuilder Worker Janitor runs this deletion. See Autobuilder Worker Janitor.

3.3.2 Autobuilder Clone Cache

Cloning repositories from scratch each time they are required was slow on the Autobuilder. We therefore have a stash of commonly used repositories pre-cloned on the Workers. Data is fetched from these during clones first, then “topped up” with later revisions from any upstream when necessary. The cache is maintained by the Autobuilder Worker Janitor. See Autobuilder Worker Janitor.

3.3.3 Autobuilder Worker Janitor

This is a process running on each Worker that performs two basic operations, including background file deletion at IO idle (see Autobuilder Target Execution Overview: Run clobberdir) and maintenance of a cache of cloned repositories to improve the speed the system can checkout repositories.

3.3.4 Shared DL_DIR

The Workers are all connected over NFS which allows DL_DIR to be shared between them. This reduces network accesses from the system and allows the build to be sped up. Usage of the directory within the build system is designed to be able to be shared over NFS.

3.3.5 Shared SSTATE_DIR

The Workers are all connected over NFS which allows the sstate directory to be shared between them. This means once a Worker has built an artifact, all the others can benefit from it. Usage of the directory within the directory is designed for sharing over NFS.

3.3.6 Resulttool

All of the different tests run as part of the build generate output into testresults.json files. This allows us to determine which tests ran in a given build and their status. Additional information, such as failure logs or the time taken to run the tests, may also be included.

Resulttool is part of OpenEmbedded-Core and is used to manipulate these json results files. It has the ability to merge files together, display reports of the test results and compare different result files.

For details, see https://wiki.yoctoproject.org/wiki/Resulttool.

3.4 run-config Target Execution

The scripts/run-config execution is where most of the work within the Autobuilder happens. It runs through a number of steps; the first are general setup steps that are run once and include:

  1. Set up any buildtools-tarball if configured.

  2. Call “buildhistory-init” if buildhistory is configured.

For each step that is configured in config.json, it will perform the following:

  1. Add any layers that are specified using the bitbake-layers add-layer command (logging as stepXa)

  2. Call the scripts/setup-config script to generate the necessary auto.conf configuration file for the build

  3. Run the bitbake BBTARGETS command (logging as stepXb)

  4. Run the bitbake SANITYTARGETS command (logging as stepXc)

  5. Run the EXTRACMDS command, which are run within the BitBake build environment (logging as stepXd)

  6. Run the EXTRAPLAINCMDS command(s), which are run outside the BitBake build environment (logging as stepXd)

  7. Remove any layers added in step 1 using the bitbake-layers remove-layer command (logging as stepXa)

Once the execution steps above complete, run-config executes a set of post-build steps, including:

  1. Call scripts/publish-artifacts to collect any output which is to be saved from the build.

  2. Call scripts/collect-results to collect any test results to be saved from the build.

  3. Call scripts/upload-error-reports to send any error reports generated to the remote server.

  4. Cleanup the build directory using clobberdir if the build was successful, else rename it to “build-renamed” for potential future debugging.

3.5 Deploying Yocto Autobuilder

The most up to date information about how to setup and deploy your own Autobuilder can be found in README.md in the yocto-autobuilder2 repository.

We hope that people can use the yocto-autobuilder2 code directly but it is inevitable that users will end up needing to heavily customise the yocto-autobuilder-helper repository, particularly the config.json file as they will want to define their own test matrix.

The Autobuilder supports wo customization options:

  • variable substitution

  • overlaying configuration files

The standard config.json minimally attempts to allow substitution of the paths. The Helper script repository includes a local-example.json file to show how you could override these from a separate configuration file. Pass the following into the environment of the Autobuilder:

$ ABHELPER_JSON="config.json local-example.json"

As another example, you could also pass the following into the environment:

$ ABHELPER_JSON="config.json /some/location/local.json"

One issue users often run into is validation of the config.json files. A tip for minimizing issues from invalid json files is to use a Git pre-commit-hook.sh script to verify the JSON file before committing it. Create a symbolic link as follows:

$ ln -s ../../scripts/pre-commit-hook.sh .git/hooks/pre-commit