Plugins

The OpenStack ecosystem is wide and deep, and only growing more so every day. The value of DevStack is that it’s simple enough to understand what it’s doing clearly. And yet we’d like to support as much of the OpenStack Ecosystem as possible. We do that with plugins.

DevStack plugins are bits of bash code that live outside the DevStack tree. They are called through a strong contract, so these plugins can be sure that they will continue to work in the future as DevStack evolves.

Prerequisites

If you are planning to create a plugin that is going to host a service in the service catalog (that is, your plugin will use the command get_or_create_service) please make sure that you apply to the service types authority to reserve a valid service-type. This will help to make sure that all deployments of your service use the same service-type.

Plugin Interface

DevStack supports a standard mechanism for including plugins from external repositories. The plugin interface assumes the following:

An external git repository that includes a devstack/ top level directory. Inside this directory there can be 3 files.

  • override-defaults - a file containing global variables that will be sourced before the lib/* files. This allows the plugin to override the defaults that are otherwise set in the lib/* files.

    For example, override-defaults may export CINDER_ENABLED_BACKENDS to include the plugin-specific storage backend and thus be able to override the default lvm only storage backend for Cinder.

  • settings - a file containing global variables that will be sourced very early in the process. This is helpful if other plugins might depend on this one, and need access to global variables to do their work.

    Your settings should include any enable_service lines required by your plugin. This is especially important if you are kicking off services using run_process as it only works with enabled services.

    Be careful to allow users to override global-variables for customizing their environment. Usually it is best to provide a default value only if the variable is unset or empty; e.g. in bash syntax FOO=${FOO:-default}.

    The file should include a define_plugin line to indicate the plugin’s name, which is the name that should be used by users on “enable_plugin” lines. It should generally be the last component of the git repo path (e.g., if the plugin’s repo is openstack/foo, then the name here should be “foo”)

    define_plugin <YOUR PLUGIN>
    

    If your plugin depends on another plugin, indicate it in this file with one or more lines like the following:

    plugin_requires <YOUR PLUGIN> <OTHER PLUGIN>
    

    For a complete example, if the plugin “foo” depends on “bar”, the settings file should include:

    define_plugin foo
    plugin_requires foo bar
    

    Devstack does not currently use this dependency information, so it’s important that users continue to add enable_plugin lines in the correct order in local.conf, however adding this information allows other tools to consider dependency information when automatically generating local.conf files.

  • plugin.sh - the actual plugin. It is executed by devstack at well defined points during a stack.sh run. The plugin.sh internal structure is discussed below.

Plugins are registered by adding the following to the localrc section of local.conf.

They are added in the following format:

[[local|localrc]]
enable_plugin <NAME> <GITURL> [GITREF]
  • name - an arbitrary name. (ex: glusterfs, docker, zaqar, congress)

  • giturl - a valid git url that can be cloned

  • gitref - an optional git ref (branch / ref / tag) that will be cloned. Defaults to master.

An example would be as follows:

enable_plugin ec2-api https://opendev.org/openstack/ec2-api

plugin.sh contract

plugin.sh is a bash script that will be called at specific points during stack.sh, unstack.sh, and clean.sh. It will be called in the following way:

source $PATH/TO/plugin.sh <mode> [phase]

mode can be thought of as the major mode being called, currently one of: stack, unstack, clean. phase is used by modes which have multiple points during their run where it’s necessary to be able to execute code. All existing mode and phase points are considered strong contracts and won’t be removed without a reasonable deprecation period. Additional new mode or phase points may be added at any time if we discover we need them to support additional kinds of plugins in devstack.

The current full list of mode and phase are:

  • stack - Called by stack.sh four times for different phases of its run:

    • pre-install - Called after system (OS) setup is complete and before project source is installed.

    • install - Called after the layer 1 and 2 projects source and their dependencies have been installed.

    • post-config - Called after the layer 1 and 2 services have been configured. All configuration files for enabled services should exist at this point.

    • extra - Called near the end after layer 1 and 2 services have been started.

    • test-config - Called at the end of devstack used to configure tempest or any other test environments

  • unstack - Called by unstack.sh before other services are shut down.

  • clean - Called by clean.sh before other services are cleaned, but after unstack.sh has been called.

Example plugin

An example plugin would look something as follows.

devstack/settings:

# settings file for template
enable_service template

devstack/plugin.sh:

# plugin.sh - DevStack plugin.sh dispatch script template

function install_template {
    ...
}

function init_template {
    ...
}

function configure_template {
    ...
}

# check for service enabled
if is_service_enabled template; then

    if [[ "$1" == "stack" && "$2" == "pre-install" ]]; then
        # Set up system services
        echo_summary "Configuring system services Template"
        install_package cowsay

    elif [[ "$1" == "stack" && "$2" == "install" ]]; then
        # Perform installation of service source
        echo_summary "Installing Template"
        install_template

    elif [[ "$1" == "stack" && "$2" == "post-config" ]]; then
        # Configure after the other layer 1 and 2 services have been configured
        echo_summary "Configuring Template"
        configure_template

    elif [[ "$1" == "stack" && "$2" == "extra" ]]; then
        # Initialize and start the template service
        echo_summary "Initializing Template"
        init_template
    fi

    if [[ "$1" == "unstack" ]]; then
        # Shut down template services
        # no-op
        :
    fi

    if [[ "$1" == "clean" ]]; then
        # Remove state and transient data
        # Remember clean.sh first calls unstack.sh
        # no-op
        :
    fi
fi

Plugin Execution Order

Plugins are run after in tree services at each of the stages above. For example, if you need something to happen before Keystone starts, you should do that at the post-config phase.

Multiple plugins can be specified in your local.conf. When that happens the plugins will be executed in order at each phase. This allows plugins to conceptually depend on each other through documenting to the user the order they must be declared. A formal dependency mechanism is beyond the scope of the current work.

System Packages

Devstack based

Devstack provides a custom framework for getting packages installed at an early phase of its execution. These packages may be defined in a plugin as files that contain new-line separated lists of packages required by the plugin

Supported packaging systems include apt and yum across multiple distributions. To enable a plugin to hook into this and install package dependencies, packages may be listed at the following locations in the top-level of the plugin repository:

  • ./devstack/files/debs/$plugin_name - Packages to install when running on Ubuntu or Debian.

  • ./devstack/files/rpms/$plugin_name - Packages to install when running on Red Hat, Fedora, or CentOS.

  • ./devstack/files/rpms-suse/$plugin_name - Packages to install when running on SUSE Linux or openSUSE.

Although there a no plans to remove this method of installing packages, plugins should consider it deprecated for bindep support described below.

bindep

The bindep project has become the defacto standard for OpenStack projects to specify binary dependencies.

A plugin may provide a ./devstack/files/bindep.txt file, which will be called with the default profile to install packages. For details on the syntax, etc. see the bindep documentation.

It is also possible to use the bindep.txt of projects that are being installed from source with the -bindep flag available in install functions. For example

if use_library_from_git "diskimage-builder"; then
   GITREPO["diskimage-builder"]=$DISKIMAGE_BUILDER_REPO_URL
   GITDIR["diskimage-builder"]=$DEST/diskimage-builder
   GITBRANCH["diskimage-builder"]=$DISKIMAGE_BUILDER_REPO_REF
   git_clone_by_name "diskimage-builder"
   setup_dev_lib -bindep "diskimage-builder"
fi

will result in any packages required by the bindep.txt of the diskimage-builder project being installed. Note however that jobs that switch projects between source and released/pypi installs (e.g. with a foo-dsvm and a foo-dsvm-src test to cover both released dependencies and master versions) will have to deal with bindep.txt being unavailable without the source directory.

Using Plugins in the OpenStack Gate

For everyday use, DevStack plugins can exist in any git tree that’s accessible on the internet. However, when using DevStack plugins in the OpenStack gate, they must live in projects in OpenStack’s gerrit. This allows testing of the plugin as well as provides network isolation against upstream git repository failures (which we see often enough to be an issue).

Ideally a plugin will be included within the devstack directory of the project they are being tested. For example, the openstack/ec2-api project has its plugin support in its own tree.

However, some times a DevStack plugin might be used solely to configure a backend service that will be used by the rest of OpenStack, so there is no “project tree” per say. Good examples include: integration of back end storage (e.g. ceph or glusterfs), integration of SDN controllers (e.g. ovn, OpenDayLight), or integration of alternate RPC systems (e.g. zmq, qpid). In these cases the best practice is to build a dedicated openstack/devstack-plugin-FOO project.

Legacy project-config jobs

To enable a plugin to be used in a gate job, the following lines will be needed in your jenkins/jobs/<project>.yaml definition in project-config:

# Because we are testing a non standard project, add the
# our project repository. This makes zuul do the right
# reference magic for testing changes.
export PROJECTS="openstack/ec2-api $PROJECTS"

# note the actual url here is somewhat irrelevant because it
# caches in nodepool, however make it a valid url for
# documentation purposes.
export DEVSTACK_LOCAL_CONFIG="enable_plugin ec2-api https://opendev.org/openstack/ec2-api"

Zuul v3 jobs

See the devstack_plugins example in Migrating Zuul V2 CI jobs to V3.

See Also

For additional inspiration on devstack plugins you can check out the Plugin Registry.