Feature discovery

Table of contents

1. Feature labels
2. Feature sources
   1. CPU
   2. Custom
   3. IOMMU
   4. Kernel
   5. Memory
   6. Network
   7. PCI
   8. USB
   9. Storage
   10. System
   11. Local – user-specific features
3. Extended resources

Feature discovery in nfd-worker is performed by a set of separate modules called feature sources. Most of them are specifically responsible for certain domain of features (e.g. cpu). In addition there are two highly customizable feature sources that work accross the system.

Feature labels

Each discovered feature is advertised a label in the Kubernetes Node object. The published node labels encode a few pieces of information:

• Namespace
  • all built-in labels use feature.node.kubernetes.io
  • user-specified custom labels (custom and local feature sources)
    • feature.node.kubernetes.io and its sub-namespaces (e.g. vendor.profile.node.kubernetes.io) are allowed by default
    • additional namespaces may be enabled with the --extra-label-ns command line flag of nfd-master
• The source for each label (e.g. cpu).
• The name of the discovered feature as it appears in the underlying source, (e.g. cpuid.AESNI from cpu).
• The value of the discovered feature.

Feature label names adhere to the following pattern:

<namespace>/<source name>-<feature name>[.<attribute name>]

The last component (i.e. attribute-name) is optional, and only used if a feature logically has sub-hierarchy, e.g. sriov.capable and sriov.configure from the network source.

The -sources flag controls which sources to use for discovery.

Note: Consecutive runs of nfd-worker will update the labels on a given node. If features are not discovered on a consecutive run, the corresponding label will be removed. This includes any restrictions placed on the consecutive run, such as restricting discovered features with the -label-whitelist option.

Feature sources

CPU

The cpu feature source supports the following labels:

The (sub-)set of CPUID attributes to publish is configurable via the attributeBlacklist and attributeWhitelist cpuid options of the cpu source. If whitelist is specified, only whitelisted attributes will be published. With blacklist, only blacklisted attributes are filtered out. attributeWhitelist has priority over attributeBlacklist. For examples and more information about configurability, see configuration. By default, the following CPUID flags have been blacklisted: BMI1, BMI2, CLMUL, CMOV, CX16, ERMS, F16C, HTT, LZCNT, MMX, MMXEXT, NX, POPCNT, RDRAND, RDSEED, RDTSCP, SGX, SSE, SSE2, SSE3, SSE4, SSE42 and SSSE3.

NOTE The cpuid features advertise supported CPU capabilities, that is, a capability might be supported but not enabled.

X86 CPUID attributes (partial list)

See the full list in github.com/klauspost/cpuid.

Arm CPUID attribute (partial list)

Arm64 CPUID attribute (partial list)

Custom

The Custom feature source allows the user to define features based on a mix of predefined rules. A rule is provided input witch affects its process of matching for a defined feature. The rules are specified in the nfd-worker configuration file. See configuration for instructions and examples how to set-up and manage the worker configuration.

To aid in making Custom Features clearer, we define a general and a per rule nomenclature, keeping things as consistent as possible.

Additional configuration directory

Additionally to the rules defined in the nfd-worker configuration file, the Custom feature can read more configuration files located in the /etc/kubernetes/node-feature-discovery/custom.d/ directory. This makes more dynamic and flexible configuration easier. This directory must be available inside the NFD worker container, so Volumes and VolumeMounts must be used for mounting e.g. ConfigMap(s). The example deployment manifests provide an example (commented out) for providing Custom configuration with an additional ConfigMap, mounted into the custom.d directory.

General nomenclature & definitions

Rule        :Represents a matching logic that is used to match on a feature.
Rule Input  :The input a Rule is provided. This determines how a Rule performs the match operation.
Matcher     :A composition of Rules, each Matcher may be composed of at most one instance of each Rule.

Custom features format (using the nomenclature defined above)

Rules are specified under sources.custom in the nfd-worker configuration file.

sources:
  custom:
  - name: <feature name>
    value: <optional feature value, defaults to "true">
    matchOn:
    - <Rule-1>: <Rule-1 Input>
      [<Rule-2>: <Rule-2 Input>]
    - <Matcher-2>
    - ...
    - ...
    - <Matcher-N>
  - <custom feature 2>
  - ...
  - ...
  - <custom feature M>

Matching process

Specifying Rules to match on a feature is done by providing a list of Matchers. Each Matcher contains one or more Rules.

Logical OR is performed between Matchers and logical AND is performed between Rules of a given Matcher.

Rules

pciid rule

Nomenclature

Attribute   :A PCI attribute.
Element     :An identifier of the PCI attribute.

The PciId Rule allows matching the PCI devices in the system on the following Attributes: class,vendor and device. A list of Elements is provided for each Attribute.

Format

pciId :
  class: [<class id>, ...]
  vendor: [<vendor id>,  ...]
  device: [<device id>, ...]

Matching is done by performing a logical OR between Elements of an Attribute and logical AND between the specified Attributes for each PCI device in the system. At least one Attribute must be specified. Missing attributes will not partake in the matching process.

UsbId rule

Nomenclature

Attribute   :A USB attribute.
Element     :An identifier of the USB attribute.

The UsbId Rule allows matching the USB devices in the system on the following Attributes: class,vendor, device and serial. A list of Elements is provided for each Attribute.

Format

usbId :
  class: [<class id>, ...]
  vendor: [<vendor id>,  ...]
  device: [<device id>, ...]
  serial: [<serial>, ...]

Matching is done by performing a logical OR between Elements of an Attribute and logical AND between the specified Attributes for each USB device in the system. At least one Attribute must be specified. Missing attributes will not partake in the matching process.

LoadedKMod rule

Nomenclature

Element     :A kernel module

The LoadedKMod Rule allows matching the loaded kernel modules in the system against a provided list of Elements.

Format

loadedKMod : [<kernel module>, ...]

Matching is done by performing logical AND for each provided Element, i.e the Rule will match if all provided Elements (kernel modules) are loaded in the system.

CpuId rule

Nomenclature

Element     :A CPUID flag

The Rule allows matching the available CPUID flags in the system against a provided list of Elements.

Format

cpuId : [<CPUID flag string>, ...]

Matching is done by performing logical AND for each provided Element, i.e the Rule will match if all provided Elements (CPUID flag strings) are available in the system.

Kconfig rule

Nomenclature

Element     :A Kconfig option

The Rule allows matching the kconfig options in the system against a provided list of Elements.

Format

kConfig: [<kernel config option ('y' or 'm') or '=<value>'>, ...]

Matching is done by performing logical AND for each provided Element, i.e the Rule will match if all provided Elements (kernel config options) are enabled (y or m) or matching =<value> in the kernel.

Nodename rule

Nomenclature

Element     :A nodename regexp pattern

The Rule allows matching the node's name against a provided list of Elements.

Format

nodename: [ <nodename regexp pattern>, ... ]

Matching is done by performing logical OR for each provided Element, i.e the Rule will match if one of the provided Elements (nodename regexp pattern) matches the node's name.

Example

custom:
  - name: "my.kernel.feature"
    matchOn:
      - loadedKMod: ["kmod1", "kmod2"]
  - name: "my.pci.feature"
    matchOn:
      - pciId:
          vendor: ["15b3"]
          device: ["1014", "1017"]
  - name: "my.usb.feature"
    matchOn:
      - usbId:
          vendor: ["1d6b"]
          device: ["0003"]
          serial: ["090129a"]
  - name: "my.combined.feature"
    matchOn:
      - loadedKMod : ["vendor_kmod1", "vendor_kmod2"]
        pciId:
          vendor: ["15b3"]
          device: ["1014", "1017"]
  - name: "vendor.feature.node.kubernetes.io/accumulated.feature"
    matchOn:
      - loadedKMod : ["some_kmod1", "some_kmod2"]
      - pciId:
          vendor: ["15b3"]
          device: ["1014", "1017"]
  - name: "my.kernel.featureneedscpu"
    matchOn:
      - kConfig: ["KVM_INTEL"]
      - cpuId: ["VMX"]
  - name: "my.kernel.modulecompiler"
    matchOn:
      - kConfig: ["GCC_VERSION=100101"]
        loadedKMod: ["kmod1"]
  - name: "my.datacenter"
    value: "datacenter-1"
    matchOn:
      - nodename: [ "node-datacenter1-rack.*-server.*" ]

In the example above:

• A node would contain the label: feature.node.kubernetes.io/custom-my.kernel.feature=true if the node has kmod1 AND kmod2 kernel modules loaded.
• A node would contain the label: feature.node.kubernetes.io/custom-my.pci.feature=true if the node contains a PCI device with a PCI vendor ID of 15b3 AND PCI device ID of 1014 OR 1017.
• A node would contain the label: feature.node.kubernetes.io/custom-my.usb.feature=true if the node contains a USB device with a USB vendor ID of 1d6b AND USB device ID of 0003.
• A node would contain the label: feature.node.kubernetes.io/custom-my.combined.feature=true if vendor_kmod1 AND vendor_kmod2 kernel modules are loaded AND the node contains a PCI device with a PCI vendor ID of 15b3 AND PCI device ID of 1014 or 1017.
• A node would contain the label: vendor.feature.node.kubernetes.io/accumulated.feature=true if some_kmod1 AND some_kmod2 kernel modules are loaded OR the node contains a PCI device with a PCI vendor ID of 15b3 AND PCI device ID of 1014 OR 1017.
• A node would contain the label: feature.node.kubernetes.io/custom-my.kernel.featureneedscpu=true if KVM_INTEL kernel config is enabled AND the node CPU supports VMX virtual machine extensions
• A node would contain the label: feature.node.kubernetes.io/custom-my.kernel.modulecompiler=true if the in-tree kmod1 kernel module is loaded AND it's built with GCC_VERSION=100101.
• A node would contain the label: feature.node.kubernetes.io/my.datacenter=datacenter-1 if the node's name matches the node-datacenter1-rack.*-server.* pattern, e.g. node-datacenter1-rack2-server42

Statically defined features

Some feature labels which are common and generic are defined statically in the custom feature source. A user may add additional Matchers to these feature labels by defining them in the nfd-worker configuration file.

IOMMU

The iommu feature source supports the following labels:

Kernel

The kernel feature source supports the following labels:

Kernel config file to use, and, the set of config options to be detected are configurable. See configuration for more information.

Memory

The memory feature source supports the following labels:

Network

The network feature source supports the following labels:

PCI

The pci feature source supports the following labels:

<device label> is composed of raw PCI IDs, separated by underscores. The set of fields used in <device label> is configurable, valid fields being class, vendor, device, subsystem_vendor and subsystem_device. Defaults are class and vendor. An example label using the default label fields:

feature.node.kubernetes.io/pci-1200_8086.present=true

Also the set of PCI device classes that the feature source detects is configurable. By default, device classes (0x)03, (0x)0b40 and (0x)12, i.e. GPUs, co-processors and accelerator cards are detected.

USB

The usb feature source supports the following labels:

<device label> is composed of raw USB IDs, separated by underscores. The set of fields used in <device label> is configurable, valid fields being class, vendor, device and serial. Defaults are class, vendor and device. An example label using the default label fields:

feature.node.kubernetes.io/usb-fe_1a6e_089a.present=true

See configuration for more information on NFD config.

Storage

The storage feature source supports the following labels:

System

The system feature source supports the following labels:

Local – user-specific features

NFD has a special feature source named local which is designed for getting the labels from user-specific feature detector. It provides a mechanism for users to implement custom feature sources in a pluggable way, without modifying nfd source code or Docker images. The local feature source can be used to advertise new user-specific features, and, for overriding labels created by the other feature sources.

The local feature source gets its labels by two different ways:

• It tries to execute files found under /etc/kubernetes/node-feature-discovery/source.d/ directory. The hook files must be executable and they are supposed to print all discovered features in stdout, one per line. With ELF binaries static linking is recommended as the selection of system libraries available in the NFD release image is very limited. Other runtimes currently supported by the NFD stock image are bash and perl.
• It reads files found under /etc/kubernetes/node-feature-discovery/features.d/ directory. The file content is expected to be similar to the hook output (described above).

NOTE: The minimal image variant only supports running statically linked binaries.

These directories must be available inside the Docker image so Volumes and VolumeMounts must be used if standard NFD images are used. The given template files mount by default the source.d and the features.d directories respectively from /etc/kubernetes/node-feature-discovery/source.d/ and /etc/kubernetes/node-feature-discovery/features.d/ from the host. You should update them to match your needs.

In both cases, the labels can be binary or non binary, using either <name> or <name>=<value> format.

Unlike the other feature sources, the name of the file, instead of the name of the feature source (that would be local in this case), is used as a prefix in the label name, normally. However, if the <name> of the label starts with a slash (/) it is used as the label name as is, without any additional prefix. This makes it possible for the user to fully control the feature label names, e.g. for overriding labels created by other feature sources.

You can also override the default namespace of your labels using this format: <namespace>/<name>[=<value>]. If using something else than [<sub-ns>.]feature.node.kubernetes.io, you must whitelist your namespace using the -extra-label-ns option on the master. In this case, the name of the file will not be added to the label name. For example, if you want to add the label my.namespace.org/my-label=value, your hook output or file must contains my.namespace.org/my-label=value and you must add -extra-label-ns=my.namespace.org on the master command line.

stderr output of the hooks is propagated to NFD log so it can be used for debugging and logging.

Injecting labels from other pods

One use case for the hooks and/or feature files is detecting features in other Pods outside NFD, e.g. in Kubernetes device plugins. It is possible to mount the source.d and/or features.d directories common with the NFD Pod and deploy the custom hooks/features there. NFD will periodically scan the directories and run any hooks and read any feature files it finds. The default deployments contain hostPath mounts for sources.d and features.d directories. By using the same mounts in the secondary Pod (e.g. device plugin) you have created a shared area for delivering hooks and feature files to NFD.

A hook example

User has a shell script /etc/kubernetes/node-feature-discovery/source.d/my-source which has the following stdout output:

MY_FEATURE_1
MY_FEATURE_2=myvalue
/override_source-OVERRIDE_BOOL
/override_source-OVERRIDE_VALUE=123
override.namespace/value=456

which, in turn, will translate into the following node labels:

feature.node.kubernetes.io/my-source-MY_FEATURE_1=true
feature.node.kubernetes.io/my-source-MY_FEATURE_2=myvalue
feature.node.kubernetes.io/override_source-OVERRIDE_BOOL=true
feature.node.kubernetes.io/override_source-OVERRIDE_VALUE=123
override.namespace/value=456

A file example

User has a file /etc/kubernetes/node-feature-discovery/features.d/my-source which contains the following lines:

MY_FEATURE_1
MY_FEATURE_2=myvalue
/override_source-OVERRIDE_BOOL
/override_source-OVERRIDE_VALUE=123
override.namespace/value=456

which, in turn, will translate into the following node labels:

feature.node.kubernetes.io/my-source-MY_FEATURE_1=true
feature.node.kubernetes.io/my-source-MY_FEATURE_2=myvalue
feature.node.kubernetes.io/override_source-OVERRIDE_BOOL=true
feature.node.kubernetes.io/override_source-OVERRIDE_VALUE=123
override.namespace/value=456

NFD tries to run any regular files found from the hooks directory. Any additional data files your hook might need (e.g. a configuration file) should be placed in a separate directory in order to avoid NFD unnecessarily trying to execute these. You can use a subdirectory under the hooks directory, for example /etc/kubernetes/node-feature-discovery/source.d/conf/.

NOTE! NFD will blindly run any executables placed/mounted in the hooks directory. It is the user's responsibility to review the hooks for e.g. possible security implications.

NOTE! Be careful when creating and/or updating hook or feature files while NFD is running. In order to avoid race conditions you should write into a temporary file (outside the source.d and features.d directories), and, atomically create/update the original file by doing a filesystem move operation.

Extended resources

This feature is experimental and by no means a replacement for the usage of device plugins.

Labels which have integer values, can be promoted to Kubernetes extended resources by listing them to the master -resource-labels command line flag. These labels won't then show in the node label section, they will appear only as extended resources.

An example use-case for the extended resources could be based on a hook which creates a label for the node SGX EPC memory section size. By giving the name of that label in the -resource-labels flag, that value will then turn into an extended resource of the node, allowing PODs to request that resource and the Kubernetes scheduler to schedule such PODs to only those nodes which have a sufficient capacity of said resource left.

Similar to labels, the default namespace feature.node.kubernetes.io is automatically prefixed to the extended resource, if the promoted label doesn't have a namespace.

Example usage of the command line arguments, using a new namespace: nfd-master -resource-labels=my_source-my.feature,sgx.some.ns/epc -extra-label-ns=sgx.some.ns

The above would result in following extended resources provided that related labels exist:

  sgx.some.ns/epc: <label value>
  feature.node.kubernetes.io/my_source-my.feature: <label value>