Jenkins configuration

This page explains common Jenkins configuration used in GROMACS builds. You may want to first read Jenkins scripts (releng Python module) to understand how the actual builds are done.

Job configuration for freestyle projects

Configuration for Jenkins projects that use the releng scripts are described here. The description in this section applies directly to freestyle (non-pipeline) builds. Pipeline builds also apply the same principles, but similarities and differences are described in the next section.

SCM checkout configuration

  • Jenkins SCM configuration should be used to check out the repository from where the build is triggered as a subdirectory of the workspace, with the same name as the repository (this creates the layout described in Jenkins scripts (releng Python module)). Using the triggering repository is necessary for the Git Plugin to show reasonable change lists for the builds etc., although the build in reality always starts from the releng repository.
  • The build script always needs to check out the releng repository if it did not trigger the build, and start the build from there.
  • The releng script will check out remaining repositories if necessary.

Build parameters and environment variables

Various *_REFSPEC environment variables (see Input environment variables) need to be set in one way or another. The suggested approach is to use build parameters as below:

To create a build that allows both intuitive parameterized builds with given refspecs and Gerrit Trigger builds, the following configuration is recommended:

  • Use GROMACS_REFSPEC, RELENG_REFSPEC, and REGRESSIONTESTS_REFSPEC (if needed) as build parameters, with refs/heads/master (or another branch ref) as the default. With pipeline builds, it is possible to also set GROMACS_REFSPEC and REGRESSIONTESTS_REFSPEC to auto as the default.

  • Use “Prepare environment for the run” and the following Groovy script:

    if (!binding.variables.containsKey('GERRIT_PROJECT')) {
      return [CHECKOUT_PROJECT: 'gromacs', CHECKOUT_REFSPEC: GROMACS_REFSPEC]
    } else {
      return [CHECKOUT_PROJECT: GERRIT_PROJECT, CHECKOUT_REFSPEC: GERRIT_REFSPEC]
    }
    
  • Configure all SCM checkout behaviors to use CHECKOUT_PROJECT and CHECKOUT_REFSPEC.

To create a build that works as expected in all corner cases when triggered from a pipeline job, the following configuration is recommended:

  • Create additional string parameters GROMACS_HASH, RELENG_HASH, and REGRESSIONTESTS_HASH with empty default values.

  • Create a string parameter CHECKOUT_PROJECT, with the default value gromacs (or another repository that you want to see in Changes section for manually triggered builds).

  • Use the following Groovy script for injecting environment variables:

    return [CHECKOUT_REFSPEC: binding.variables."${CHECKOUT_PROJECT.toUpperCase()}_REFSPEC"]
    

    If you also need to support directly triggering the build with Gerrit Trigger, you need a slightly more complicated script, but in most cases, it should be the pipeline job that is triggered with Gerrit Trigger.

In SCM poll jobs it is possible to simply set the various environment variables to static values using a properties file in “Prepare environment for the run” (CHECKOUT_PROJECT and the various *_REFSPEC variables). Note that the SCM checkout behavior cannot use CHECKOUT_PROJECT in the git address, because the injected variables are not available for SCM polling.

Build steps

Builds that call run_build() should use the following post-build steps:

  • The job should check the console output for the string “FAILED” and mark the build unstable if this is found.
  • The job should use logs/unsuccessful-reason.log as the “Unsuccessful Message File” for the Gerrit Trigger plugin. TODO: How to best handle this for matrix builds (or other types of multi-configuration builds)
  • The job should archive all .log files from logs/. Note that the build should be configured not to fail if there is nothing to archive if all the logs are conditionally produced.
  • The job can check various log files under logs/category/ for warnings; the general design is that all logs from a certain category are checked using the same warning parser.

The build script in Jenkins will look something like this:

import os
import shlex
import subprocess
import sys

# For builds not triggered by Gerrit Trigger, the conditional is not
# necessary.
if os.environ['CHECKOUT_PROJECT'] != 'releng':
    if not os.path.isdir('releng'):
        os.makedirs('releng')
    os.chdir('releng')
    subprocess.check_call(['git', 'init'])
    subprocess.check_call(['git', 'fetch', 'ssh://jenkins@gerrit.gromacs.org/releng.git', os.environ['RELENG_REFSPEC']])
    subprocess.check_call(['git', 'checkout', '-qf', 'FETCH_HEAD'])
    subprocess.check_call(['git', 'clean', '-ffdxq'])
    subprocess.check_call(['git', 'gc'])
    os.chdir('..')

sys.path.append(os.path.abspath('releng'))
import releng

# For non-matrix builds, opts can be a hard-coded list (or possibly None).
opts = shlex.split(os.environ['OPTIONS'])
releng.run_build('gromacs', releng.JobType.GERRIT, opts)

The script checks out the releng repository to a releng/ subdirectory of the workspace if not already checked out, imports the releng package and runs run_build() with arguments identifying which build script to run, and options that affect how the build is done. shlex.split() is necessary to be able to pass quoted arguments with spaces to options (not currently used).

Matrix builds are nowadays triggered through a pipeline build that chooses the build hosts dynamically inside the releng Python scripts. The scripts still support using with a host= or a label= option in the options to select the host, and that option is automatically ignored by run_build().

run_build() will first check out the gromacs repository to a gromacs/ subdirectory of the workspace, and then execute a script from gromacs/admin/builds/, selected based on the first argument. If necessary, it will also check out the regression tests repository. If the script exits with a non-zero exit code, the build fails.

Job configuration for pipeline builds

For pipeline job configuration, the same principles apply as for freestyle projects, but much more is handled in the pipeline Groovy script instead of in job configuration.

  • SCM checkout as described above is handled by utils.checkoutDefaultProject(), called from the beginning of each pipeline script. Jenkins only needs to checkout the releng repository to load the Groovy script (see the bootstrap script below).

  • Build parameters for GROMACS_REFSPEC, RELENG_REFSPEC, and REGRESSIONTESTS_REFSPEC (if needed) should be added as for freestyle projects. There is no need to deal with CHECKOUT_PROJECT or with environment variables explicitly (the environment injection plugin does not work with pipeline builds, either). All processing of the parameters is done by utils.initBuildRevisions() at the start of each Groovy script.

    For GROMACS_REFSPEC and REGRESSIONTESTS_REFSPEC, it is possible to use auto as the default value to create jobs that can be triggered for multiple branches from Gerrit or manually by specifying only one refspec.

  • CHECKOUT_PROJECT must not be used as a build parameter (would currently confuse the Python scripts launched from Groovy).

  • *_HASH parameters can be used as with freestyle projects. If not set, they are computed at the beginning in utils.initBuildRevisions().

  • In freestyle jobs, build status handling required scanning the console log and using unsuccessful-reason.log. In pipeline builds, this is handled inside utils.groovy whenever Python scripts are invoked, and uses return status of Python and a .json file created by the Python code.

Pipeline builds use a bootstrapping script like this:

def script
node('pipeline-general') {
    def checkout_refspec = params.RELENG_REFSPEC
    if (params.GERRIT_PROJECT == 'releng') {
        checkout_refspec = params.GERRIT_REFSPEC
    }
    sh """\
        set -e
        mkdir -p releng
        cd releng
        git init
        git fetch ssh://jenkins@gerrit.gromacs.org/releng.git ${checkout_refspec}
        git checkout -qf FETCH_HEAD
        git clean -ffdxq
        git gc
        """.stripIndent()
    script = load 'releng/workflow/<pipeline-name>.groovy'
    <possible additional calls as needed by the pipeline>
}
script.doBuild(<possible additional parameters>)

where expressions in angle brackets depend on the pipeline. For pipeline that are never triggered by Gerrit Trigger from releng, the part referencing GERRIT_PROJECT and GERRIT_REFSPEC can be omitted.

Jenkins plugins

The following Jenkins plugins are used in GROMACS builds:

TODO

Build agent labels

The following labels on the Jenkins build agents are currently used to allocate builds to agents:

pipeline-master
Used to run general steps in pipeline jobs that do not do any lengthy processing (except for source code checkouts). These could in principle run anywhere, but limiting them to a subset of the nodes reduces the number of workspaces used. This reduces disk space use, and each time a new workspace is created, the initial checkout takes quite a bit of time.
clang-static-analyzer-X.Y
Used to run clang static analysis builds. The build is dynamically allocated using a version-specific label, based on what is specified in the clang-analyzer.py build script in the source repository.
cppcheck
Used to run cppcheck builds for release-2018 and earlier. For now, there is no version specification: all used versions of cppcheck must be installed on each node.
doxygen
Used to run documentation builds. In addition to Doxygen, also other tools needed by the documentation build (Sphinx, Latex) need to be installed here. Also the source packaging builds use this label, since they need Sphinx.
linux
Used for regression test packaging builds to get a uniform enough environment.
windows
Should not be currently used, but has been used to restrict Unix-specific things in pipelines to not run on Windows agents.

In other cases, agents are explicitly assigned to a node. Multi-configuration builds are currently assigned to nodes based on information in agents.py, not on labels configured in Jenkins.