mirror of
https://github.com/NVIDIA/nvidia-container-toolkit
synced 2024-11-29 16:02:10 +00:00
bc4bfb94a2
Signed-off-by: Evan Lezar <elezar@nvidia.com>
273 lines
11 KiB
Go
273 lines
11 KiB
Go
// Package cdi has the primary purpose of providing an API for
|
|
// interacting with CDI and consuming CDI devices.
|
|
//
|
|
// For more information about Container Device Interface, please refer to
|
|
// https://github.com/container-orchestrated-devices/container-device-interface
|
|
//
|
|
// Container Device Interface
|
|
//
|
|
// Container Device Interface, or CDI for short, provides comprehensive
|
|
// third party device support for container runtimes. CDI uses vendor
|
|
// provided specification files, CDI Specs for short, to describe how a
|
|
// container's runtime environment should be modified when one or more
|
|
// of the vendor-specific devices is injected into the container. Beyond
|
|
// describing the low level platform-specific details of how to gain
|
|
// basic access to a device, CDI Specs allow more fine-grained device
|
|
// initialization, and the automatic injection of any necessary vendor-
|
|
// or device-specific software that might be required for a container
|
|
// to use a device or take full advantage of it.
|
|
//
|
|
// In the CDI device model containers request access to a device using
|
|
// fully qualified device names, qualified names for short, consisting of
|
|
// a vendor identifier, a device class and a device name or identifier.
|
|
// These pieces of information together uniquely identify a device among
|
|
// all device vendors, classes and device instances.
|
|
//
|
|
// This package implements an API for easy consumption of CDI. The API
|
|
// implements discovery, loading and caching of CDI Specs and injection
|
|
// of CDI devices into containers. This is the most common functionality
|
|
// the vast majority of CDI consumers need. The API should be usable both
|
|
// by OCI runtime clients and runtime implementations.
|
|
//
|
|
// CDI Registry
|
|
//
|
|
// The primary interface to interact with CDI devices is the Registry. It
|
|
// is essentially a cache of all Specs and devices discovered in standard
|
|
// CDI directories on the host. The registry has two main functionality,
|
|
// injecting devices into an OCI Spec and refreshing the cache of CDI
|
|
// Specs and devices.
|
|
//
|
|
// Device Injection
|
|
//
|
|
// Using the Registry one can inject CDI devices into a container with code
|
|
// similar to the following snippet:
|
|
//
|
|
// import (
|
|
// "fmt"
|
|
// "strings"
|
|
//
|
|
// log "github.com/sirupsen/logrus"
|
|
//
|
|
// "github.com/container-orchestrated-devices/container-device-interface/pkg/cdi"
|
|
// oci "github.com/opencontainers/runtime-spec/specs-go"
|
|
// )
|
|
//
|
|
// func injectCDIDevices(spec *oci.Spec, devices []string) error {
|
|
// log.Debug("pristine OCI Spec: %s", dumpSpec(spec))
|
|
//
|
|
// unresolved, err := cdi.GetRegistry().InjectDevices(spec, devices)
|
|
// if err != nil {
|
|
// return fmt.Errorf("CDI device injection failed: %w", err)
|
|
// }
|
|
//
|
|
// log.Debug("CDI-updated OCI Spec: %s", dumpSpec(spec))
|
|
// return nil
|
|
// }
|
|
//
|
|
// Cache Refresh
|
|
//
|
|
// By default the CDI Spec cache monitors the configured Spec directories
|
|
// and automatically refreshes itself when necessary. This behavior can be
|
|
// disabled using the WithAutoRefresh(false) option.
|
|
//
|
|
// Failure to set up monitoring for a Spec directory causes the directory to
|
|
// get ignored and an error to be recorded among the Spec directory errors.
|
|
// These errors can be queried using the GetSpecDirErrors() function. If the
|
|
// error condition is transient, for instance a missing directory which later
|
|
// gets created, the corresponding error will be removed once the condition
|
|
// is over.
|
|
//
|
|
// With auto-refresh enabled injecting any CDI devices can be done without
|
|
// an explicit call to Refresh(), using a code snippet similar to the
|
|
// following:
|
|
//
|
|
// In a runtime implementation one typically wants to make sure the
|
|
// CDI Spec cache is up to date before performing device injection.
|
|
// A code snippet similar to the following accmplishes that:
|
|
//
|
|
// import (
|
|
// "fmt"
|
|
// "strings"
|
|
//
|
|
// log "github.com/sirupsen/logrus"
|
|
//
|
|
// "github.com/container-orchestrated-devices/container-device-interface/pkg/cdi"
|
|
// oci "github.com/opencontainers/runtime-spec/specs-go"
|
|
// )
|
|
//
|
|
// func injectCDIDevices(spec *oci.Spec, devices []string) error {
|
|
// registry := cdi.GetRegistry()
|
|
//
|
|
// if err := registry.Refresh(); err != nil {
|
|
// // Note:
|
|
// // It is up to the implementation to decide whether
|
|
// // to abort injection on errors. A failed Refresh()
|
|
// // does not necessarily render the registry unusable.
|
|
// // For instance, a parse error in a Spec file for
|
|
// // vendor A does not have any effect on devices of
|
|
// // vendor B...
|
|
// log.Warnf("pre-injection Refresh() failed: %v", err)
|
|
// }
|
|
//
|
|
// log.Debug("pristine OCI Spec: %s", dumpSpec(spec))
|
|
//
|
|
// unresolved, err := registry.InjectDevices(spec, devices)
|
|
// if err != nil {
|
|
// return fmt.Errorf("CDI device injection failed: %w", err)
|
|
// }
|
|
//
|
|
// log.Debug("CDI-updated OCI Spec: %s", dumpSpec(spec))
|
|
// return nil
|
|
// }
|
|
//
|
|
// Generated Spec Files, Multiple Directories, Device Precedence
|
|
//
|
|
// It is often necessary to generate Spec files dynamically. On some
|
|
// systems the available or usable set of CDI devices might change
|
|
// dynamically which then needs to be reflected in CDI Specs. For
|
|
// some device classes it makes sense to enumerate the available
|
|
// devices at every boot and generate Spec file entries for each
|
|
// device found. Some CDI devices might need special client- or
|
|
// request-specific configuration which can only be fulfilled by
|
|
// dynamically generated client-specific entries in transient Spec
|
|
// files.
|
|
//
|
|
// CDI can collect Spec files from multiple directories. Spec files are
|
|
// automatically assigned priorities according to which directory they
|
|
// were loaded from. The later a directory occurs in the list of CDI
|
|
// directories to scan, the higher priority Spec files loaded from that
|
|
// directory are assigned to. When two or more Spec files define the
|
|
// same device, conflict is resolved by chosing the definition from the
|
|
// Spec file with the highest priority.
|
|
//
|
|
// The default CDI directory configuration is chosen to encourage
|
|
// separating dynamically generated CDI Spec files from static ones.
|
|
// The default directories are '/etc/cdi' and '/var/run/cdi'. By putting
|
|
// dynamically generated Spec files under '/var/run/cdi', those take
|
|
// precedence over static ones in '/etc/cdi'. With this scheme, static
|
|
// Spec files, typically installed by distro-specific packages, go into
|
|
// '/etc/cdi' while all the dynamically generated Spec files, transient
|
|
// or other, go into '/var/run/cdi'.
|
|
//
|
|
// Spec File Generation
|
|
//
|
|
// CDI offers two functions for writing and removing dynamically generated
|
|
// Specs from CDI Spec directories. These functions, WriteSpec() and
|
|
// RemoveSpec() implicitly follow the principle of separating dynamic Specs
|
|
// from the rest and therefore always write to and remove Specs from the
|
|
// last configured directory.
|
|
//
|
|
// Corresponding functions are also provided for generating names for Spec
|
|
// files. These functions follow a simple naming convention to ensure that
|
|
// multiple entities generating Spec files simultaneously on the same host
|
|
// do not end up using conflicting Spec file names. GenerateSpecName(),
|
|
// GenerateNameForSpec(), GenerateTransientSpecName(), and
|
|
// GenerateTransientNameForSpec() all generate names which can be passed
|
|
// as such to WriteSpec() and subsequently to RemoveSpec().
|
|
//
|
|
// Generating a Spec file for a vendor/device class can be done with a
|
|
// code snippet similar to the following:
|
|
//
|
|
// import (
|
|
// "fmt"
|
|
// ...
|
|
// "github.com/container-orchestrated-devices/container-device-interface/specs-go"
|
|
// "github.com/container-orchestrated-devices/container-device-interface/pkg/cdi"
|
|
// )
|
|
//
|
|
// func generateDeviceSpecs() error {
|
|
// registry := cdi.GetRegistry()
|
|
// spec := &specs.Spec{
|
|
// Version: specs.CurrentVersion,
|
|
// Kind: vendor+"/"+class,
|
|
// }
|
|
//
|
|
// for _, dev := range enumerateDevices() {
|
|
// spec.Devices = append(spec.Devices, specs.Device{
|
|
// Name: dev.Name,
|
|
// ContainerEdits: getContainerEditsForDevice(dev),
|
|
// })
|
|
// }
|
|
//
|
|
// specName, err := cdi.GenerateNameForSpec(spec)
|
|
// if err != nil {
|
|
// return fmt.Errorf("failed to generate Spec name: %w", err)
|
|
// }
|
|
//
|
|
// return registry.SpecDB().WriteSpec(spec, specName)
|
|
// }
|
|
//
|
|
// Similary, generating and later cleaning up transient Spec files can be
|
|
// done with code fragments similar to the following. These transient Spec
|
|
// files are temporary Spec files with container-specific parametrization.
|
|
// They are typically created before the associated container is created
|
|
// and removed once that container is removed.
|
|
//
|
|
// import (
|
|
// "fmt"
|
|
// ...
|
|
// "github.com/container-orchestrated-devices/container-device-interface/specs-go"
|
|
// "github.com/container-orchestrated-devices/container-device-interface/pkg/cdi"
|
|
// )
|
|
//
|
|
// func generateTransientSpec(ctr Container) error {
|
|
// registry := cdi.GetRegistry()
|
|
// devices := getContainerDevs(ctr, vendor, class)
|
|
// spec := &specs.Spec{
|
|
// Version: specs.CurrentVersion,
|
|
// Kind: vendor+"/"+class,
|
|
// }
|
|
//
|
|
// for _, dev := range devices {
|
|
// spec.Devices = append(spec.Devices, specs.Device{
|
|
// // the generated name needs to be unique within the
|
|
// // vendor/class domain on the host/node.
|
|
// Name: generateUniqueDevName(dev, ctr),
|
|
// ContainerEdits: getEditsForContainer(dev),
|
|
// })
|
|
// }
|
|
//
|
|
// // transientID is expected to guarantee that the Spec file name
|
|
// // generated using <vendor, class, transientID> is unique within
|
|
// // the host/node. If more than one device is allocated with the
|
|
// // same vendor/class domain, either all generated Spec entries
|
|
// // should go to a single Spec file (like in this sample snippet),
|
|
// // or transientID should be unique for each generated Spec file.
|
|
// transientID := getSomeSufficientlyUniqueIDForContainer(ctr)
|
|
// specName, err := cdi.GenerateNameForTransientSpec(vendor, class, transientID)
|
|
// if err != nil {
|
|
// return fmt.Errorf("failed to generate Spec name: %w", err)
|
|
// }
|
|
//
|
|
// return registry.SpecDB().WriteSpec(spec, specName)
|
|
// }
|
|
//
|
|
// func removeTransientSpec(ctr Container) error {
|
|
// registry := cdi.GetRegistry()
|
|
// transientID := getSomeSufficientlyUniqueIDForContainer(ctr)
|
|
// specName := cdi.GenerateNameForTransientSpec(vendor, class, transientID)
|
|
//
|
|
// return registry.SpecDB().RemoveSpec(specName)
|
|
// }
|
|
//
|
|
// CDI Spec Validation
|
|
//
|
|
// This package performs both syntactic and semantic validation of CDI
|
|
// Spec file data when a Spec file is loaded via the registry or using
|
|
// the ReadSpec API function. As part of the semantic verification, the
|
|
// Spec file is verified against the CDI Spec JSON validation schema.
|
|
//
|
|
// If a valid externally provided JSON validation schema is found in
|
|
// the filesystem at /etc/cdi/schema/schema.json it is loaded and used
|
|
// as the default validation schema. If such a file is not found or
|
|
// fails to load, an embedded no-op schema is used.
|
|
//
|
|
// The used validation schema can also be changed programmatically using
|
|
// the SetSchema API convenience function. This function also accepts
|
|
// the special "builtin" (BuiltinSchemaName) and "none" (NoneSchemaName)
|
|
// schema names which switch the used schema to the in-repo validation
|
|
// schema embedded into the binary or the now default no-op schema
|
|
// correspondingly. Other names are interpreted as the path to the actual
|
|
// validation schema to load and use.
|
|
package cdi
|