Skip to content

Latest commit

 

History

History
369 lines (280 loc) · 13.9 KB

File metadata and controls

369 lines (280 loc) · 13.9 KB

Instructions for Device Plugin Development and Maintenance

Table of Contents

Day-to-day Development How to's

Get the Source Code

With git installed on the system, just clone the repository:

$ export INTEL_DEVICE_PLUGINS_SRC=/path/to/intel-device-plugins-for-kubernetes
$ git clone https://github.com/intel/intel-device-plugins-for-kubernetes ${INTEL_DEVICE_PLUGINS_SRC}

Build and Run Plugin Binaries

With go development environment installed on the system, build the plugin:

$ cd ${INTEL_DEVICE_PLUGINS_SRC}
$ make <plugin-build-target>

Note: All the available plugin build targets is roughly the output of ls ${INTEL_DEVICE_PLUGINS_SRC}/cmd.

To test the plugin binary on the development system, run as administrator:

$ sudo -E ${INTEL_DEVICE_PLUGINS_SRC}/cmd/<plugin-build-target>/<plugin-build-target>

Build Container Images

The dockerfiles are generated on the fly from .in suffixed files and .docker include-snippets which are stitched together with cpp preprocessor. You need to install cpp for that, e.g. in ubuntu it is found from build-essential (sudo apt install build-essential). Don't edit the generated dockerfiles. Edit the inputs.

The simplest way to build all the docker images, is:

$ make images

But it is very slow. You can drastically speed it up by first running once:

$ make vendor

Which brings the libraries into the builder container without downloading them again and again for each plugin.

But it is still slow. You can further speed it up by first running once:

$ make licenses

Which pre-creates the go-licenses for all plugins, instead of re-creating them for each built plugin, every time.

But it is still rather slow to build all the images, and unnecessary, if you iterate on just one. Instead, build just the one you are iterating on, example:

$ make <image-build-target>

Note: All the available image build targets is roughly the output of ls ${INTEL_DEVICE_PLUGINS_SRC}/build/docker/*.Dockerfile.

If you iterate on only one plugin and if you know what its target cmd is (see folder cmd/), you can opt to pre-create just its licenses, example:

$ make licenses/<plugin-build-target>

The container image target names in the Makefile are derived from the .Dockerfile.in suffixed filenames under folder build/docker/templates/.

Recap:

$ make vendor
$ make licenses (or just make licenses/<plugin-build-target>)
$ make <image-build-target>

Repeat the last step only, unless you change library dependencies. If you pull in new sources, start again from make vendor.

Note: The image build tool can be changed from the default docker by setting the BUILDER argument to the Makefile: make <image-build-target> BUILDER=<builder>. Supported values are docker, buildah, and podman.

Build Against a Newer Version of Kubernetes

First, you need to update module dependencies. The easiest way is to use scripts/upgrade_k8s.sh copied from a k/k issue:

Just run it inside the repo's root, e.g.

$ ${INTEL_DEVICE_PLUGINS_SRC}/scripts/upgrade_k8s.sh <k8s version>

Finally, run:

$ make generate
$ make test

and fix all new compilation issues.

Work with Intel Device Plugins Operator Modifications

There are few useful steps when working with changes to Device Plugins CRDs and controllers:

  1. Install controller-gen: GO111MODULE=on go get -u sigs.k8s.io/controller-tools/cmd/controller-gen@<release ver>, e.g, v0.4.1
  2. Generate CRD and Webhook artifacts: make generate
  3. Test local changes using envtest: make envtest
  4. Build a custom operator image: make intel-deviceplugin-operator
  5. (Un)deploy operator: kubectl [apply|delete] -k deployments/operator/default

Publish a New Version of the Intel Device Plugins Operator to operatorhub.io

Check if the fields mentioned below in the base CSV manifest file have the correct values. If not, fix them manually (operator-sdk does not support updating these fields in any other way).

  • spec.version
  • spec.replaces
  • metadata.annotations.containerImage
  • metadata.annotations.createdAT

Fork the Community Operators repo and clone it:

$ git clone https://github.com/<GitHub Username>/community-operators

Generate bundle and build bundle image:

$ make bundle TAG=0.X.Y CHANNELS=alpha DEFAULT_CHANNEL=alpha
$ make bundle-build

Push the image to a registry:

  • If pushing to the Docker hub, specify docker.io/ in front of the image name for running bundle.
  • If pushing to the local registry, put the option --use-http for running bundle.

Verify the operator deployment works OK via OLM in your development cluster:

$ operator-sdk olm install
$ kubectl create namespace testoperator
$ operator-sdk run bundle <Registry>:<Tag> -n testoperator
# do verification checks
...
# do clean up
$ operator-sdk cleanup intel-device-plugins-operator --namespace testoperator
$ kubectl delete namespace testoperator
$ operator-sdk olm uninstall

Commit files:

$ cd community-operators
$ git add operators/intel-device-plugins-operator/0.X.Y
$ git commit -am 'operators intel-device-plugins-operator (0.X.Y)' -s

Submit a PR to Community Operators repo.

Check operator page https://operatorhub.io/operator/intel-device-plugins-operator after PR is merged.

Run E2E Tests

Currently the E2E tests require having a Kubernetes cluster already configured on the nodes with the hardware required by the device plugins. Also all the container images with the executables under test must be available in the cluster. If these two conditions are satisfied, run the tests with:

$ go test -v ./test/e2e/...

In case you want to run only certain tests, e.g., QAT ones, run:

$ go test -v ./test/e2e/... -args -ginkgo.focus "QAT"

If you need to specify paths to your custom kubeconfig containing embedded authentication info then add the -kubeconfig argument:

$ go test -v ./test/e2e/... -args -kubeconfig /path/to/kubeconfig

The full list of available options can be obtained with:

$ go test ./test/e2e/... -args -help

It is also possible to run the tests which don't depend on hardware without a pre-configured Kubernetes cluster. Just make sure you have Kind installed on your host and run:

$ make test-with-kind

Run Controller Tests with a Local Control Plane

The controller-runtime library provides a package for integration testing by starting a local control plane. The package is called envtest. The operator uses this package for its integration testing.

For setting up the environment for testing, setup-envtest can be used:

$ go install sigs.k8s.io/controller-runtime/tools/setup-envtest@latest
$ setup-envtest use <K8S_VERSION>
$ KUBEBUILDER_ASSETS=$(setup-envtest use -i -p path <K8S_VERSION>) make envtest

How to Develop Simple Device Plugins

To create a simple device plugin without the hassle of developing your own gRPC server, you can use a package included in this repository called github.com/intel/intel-device-plugins-for-kubernetes/pkg/deviceplugin.

All you have to do is instantiate a deviceplugin.Manager and call its Run() method:

func main() {
    ...

    manager := dpapi.NewManager(namespace, plugin)
    manager.Run()
}

The manager's constructor accepts two parameters:

  1. namespace which is a string like "color.example.com". All your devices will be exposed under this name space, e.g. "color.example.com/yellow". Please note that one device plugin can register many such "colors". The manager will instantiate multiple gRPC servers for every registered "color".
  2. plugin which is a reference to an object implementing one mandatory interface deviceplugin.Scanner.

deviceplugin.Scanner defines one method Scan() which is called only once for every device plugin by deviceplugin.Manager in a goroutine and operates in an infinite loop. A Scan() implementation scans the host for devices and sends all found devices to a deviceplugin.Notifier instance. The deviceplugin.Notifier is implemented and provided by the deviceplugin package itself. The found devices are organized in an instance of deviceplugin.DeviceTree object. The object is filled in with its AddDevice() method:

func (dp *devicePlugin) Scan(notifier deviceplugin.Notifier) error {
    for {
        devTree := deviceplugin.NewDeviceTree()
        ...
        devTree.AddDevice("yellow", devID, deviceplugin.DeviceInfo{
            State: health,
            Nodes: []pluginapi.DeviceSpec{
                {
                    HostPath:      devPath,
                    ContainerPath: devPath,
                    Permissions:   "rw",
                },
            },
        })
        ...
        notifier.Notify(devTree)
    }
}

Optionally, your device plugin may also implement the deviceplugin.PostAllocator interface. If implemented, its method PostAllocate() modifies pluginapi.AllocateResponse responses just before they are sent to kubelet. To see an example, refer to the FPGA plugin which implements this interface to annotate its responses.

In case you want to implement the whole allocation functionality in your device plugin, you can implement the optional deviceplugin.Allocator interface. In this case PostAllocate() is not called. But if you decide in your implementation of deviceplugin.Allocator that you need to resort to the default implementation of the allocation functionality then return an error of the type deviceplugin.UseDefaultMethodError.

Logging

The framework uses klog as its logging framework. It is encouraged for plugins to also use klog to maintain uniformity in the logs and command line options.

The framework initialises klog, so further calls to klog.InitFlags() by plugins should not be necessary. This does add a number of log configuration options to your plugin, which can be viewed with the -h command line option of your plugin.

The framework tries to adhere to the Kubernetes Logging Conventions. The advise is to use the V() levels for Info() calls, as calling Info() with no set level will make configuration and filtering of logging via the command line more difficult.

The default is to not log Info() calls. This can be changed using the plugin command line -v parameter. The additional annotations prepended to log lines by 'klog' can be disabled with the -skip_headers option.

Error Conventions

The framework has a convention for producing and logging errors. Ideally plugins will also adhere to the convention.

Errors generated within the framework and plugins are instantiated with the New() and Errorf() functions of the errors package:

    return errors.New("error message")

Errors generated from outside the plugins and framework are augmented with their stack dump with code such as

    return errors.WithStack(err)

or

    return errors.Wrap(err, "some additional error message")

These errors are then logged using a default struct value format like:

    klog.Errorf("Example of an internal error death: %+v", err)

at the line where it's certain that the error cannot be passed out farther nor handled gracefully. Otherwise, they can be logged as simple values:

    klog.Warningf("Example of a warning due to an external error: %v", err)

Checklist for New Device Plugins

For new device plugins contributed to this repository, below is a checklist to get the plugin on par feature and quality wise with others:

  1. Plugin binary available in cmd/, its corresponding Dockerfile in build/docker/ and deployment Kustomization/YAMLs in deployments/.
  2. Plugin binary Go unit tests implemented and passing with >80% coverage: make test WHAT=./cmd/<plugin>.
  3. Plugin binary linter checks passing: make lint.
  4. Plugin e2e tests implemented in test/e2e/ and passing: go test -v ./test/e2e/... -args -ginkgo.focus "<plugin>".
  5. Plugin CRD API added to pkg/apis/deviceplugin/v1 and CRDs generated: make generate.
  6. Plugin CRD validation tests implemented in test/envtest/ and passing: make envtest.
  7. Plugin CRD controller implemented in pkg/controllers/ and added to the manager in cmd/operator/main.go.
  8. Plugin documentation written cmd/<plugin>/README.md and optionally end to end demos created in demo.