Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
145 changes: 145 additions & 0 deletions cmd/gpu-mutating-webhook/admission_controller.go
Original file line number Diff line number Diff line change
@@ -0,0 +1,145 @@
/**
# Copyright 2024 NVIDIA CORPORATION
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
**/

package main

import (
"encoding/json"
"errors"
"fmt"
"io/ioutil"
"log"
"net/http"

admissionv1 "k8s.io/api/admission/v1"

Check failure on line 27 in cmd/gpu-mutating-webhook/admission_controller.go

View workflow job for this annotation

GitHub Actions / golang / Build

cannot find module providing package k8s.io/api/admission/v1: import lookup disabled by -mod=vendor

Check failure on line 27 in cmd/gpu-mutating-webhook/admission_controller.go

View workflow job for this annotation

GitHub Actions / code-scanning / Analyze Go code with CodeQL

cannot find module providing package k8s.io/api/admission/v1: import lookup disabled by -mod=vendor

Check failure on line 27 in cmd/gpu-mutating-webhook/admission_controller.go

View workflow job for this annotation

GitHub Actions / golang / Unit test

cannot find module providing package k8s.io/api/admission/v1: import lookup disabled by -mod=vendor
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
"k8s.io/apimachinery/pkg/runtime"
"k8s.io/apimachinery/pkg/runtime/serializer"
)

const (
jsonContentType = `application/json`
)

var (
universalDeserializer = serializer.NewCodecFactory(runtime.NewScheme()).UniversalDeserializer()
)

type patchOperation struct {
Op string `json:"op"`
Path string `json:"path"`
Value interface{} `json:"value,omitempty"`
}

type admitFunc func(*admissionv1.AdmissionRequest) ([]patchOperation, error)

// Swati: skip nvidia-dra-driver-gpu ns as well
func isKubeNamespace(ns string) bool {
return (ns == metav1.NamespacePublic || ns == metav1.NamespaceSystem)
}

func doServeAdmitFunc(w http.ResponseWriter, r *http.Request, admit admitFunc) ([]byte, error) {
// Request validation. Only handle POST requests with a body and json content type.
if r.Method != http.MethodPost {
w.WriteHeader(http.StatusMethodNotAllowed)
return nil, fmt.Errorf("invalid method %s, only POST is allowed", r.Method)
}

body, err := ioutil.ReadAll(r.Body)
if err != nil {
w.WriteHeader(http.StatusBadRequest)
return nil, fmt.Errorf("could not read request body: %v", err)
}

if ct := r.Header.Get("Content-Type"); ct != jsonContentType {
w.WriteHeader(http.StatusBadRequest)
return nil, fmt.Errorf("unsupported content type %s, only %s is supported", ct, jsonContentType)
}

// Parse the AdmissionReview request.
var admissionReviewReq admissionv1.AdmissionReview
if _, _, err := universalDeserializer.Decode(body, nil, &admissionReviewReq); err != nil {
w.WriteHeader(http.StatusBadRequest)
return nil, fmt.Errorf("could not deserialize AdmissionReview: %v", err)
} else if admissionReviewReq.Request == nil {
w.WriteHeader(http.StatusBadRequest)
return nil, errors.New("malformed admission review: Request is nil")
}

// Build the response
admissionReviewResp := admissionv1.AdmissionReview{
TypeMeta: admissionReviewReq.TypeMeta,
Response: &admissionv1.AdmissionResponse{
UID: admissionReviewReq.Request.UID,
},
}

// Skip k8s namespaces
var patchOps []patchOperation
if !isKubeNamespace(admissionReviewReq.Request.Namespace) {
patchOps, err = admit(admissionReviewReq.Request)
}

if err != nil {
admissionReviewResp.Response.Allowed = false
admissionReviewResp.Response.Result = &metav1.Status{
Message: err.Error(),
}
} else {
patchBytes, err := json.Marshal(patchOps)
if err != nil {
w.WriteHeader(http.StatusInternalServerError)
return nil, fmt.Errorf("could not marshal JSON patch: %v", err)
}
admissionReviewResp.Response.Allowed = true
admissionReviewResp.Response.Patch = patchBytes

pt := admissionv1.PatchTypeJSONPatch
admissionReviewResp.Response.PatchType = &pt
}

respBytes, err := json.Marshal(admissionReviewResp)
if err != nil {
return nil, fmt.Errorf("could not marshal AdmissionReview response: %v", err)
}
return respBytes, nil
}

// serveAdmitFunc is a wrapper that handles HTTP, calls doServeAdmitFunc, and writes the result.
func serveAdmitFunc(w http.ResponseWriter, r *http.Request, admit admitFunc) {
log.Print("Handling webhook request ...")

respBytes, err := doServeAdmitFunc(w, r, admit)
if err != nil {
log.Printf("Error handling webhook request: %v", err)
w.WriteHeader(http.StatusInternalServerError)
_, _ = w.Write([]byte(err.Error()))
return
}

log.Print("Webhook request handled successfully")
_, writeErr := w.Write(respBytes)
if writeErr != nil {
log.Printf("Could not write response: %v", writeErr)
}
}

// admitFuncHandler converts an admitFunc into an http.Handler
func admitFuncHandler(admit admitFunc) http.Handler {
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
serveAdmitFunc(w, r, admit)
})
}
195 changes: 195 additions & 0 deletions cmd/gpu-mutating-webhook/main.go
Original file line number Diff line number Diff line change
@@ -0,0 +1,195 @@
/**
# Copyright 2025 NVIDIA CORPORATION
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
**/

package main

import (
"fmt"
"log"
"net/http"
"path/filepath"
"strings"

admissionv1 "k8s.io/api/admission/v1"
corev1 "k8s.io/api/core/v1"
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
"k8s.io/klog/v2"
)

const (
tlsDir = `/etc/webhook/tls`
tlsCertFile = `tls.crt`
tlsKeyFile = `tls.key`
gpuResourceName = "nvidia.com/gpu"
gpuClaimName = "nvidia-gpu-resourceclaim"
gpuTemplateName = "nvidia-gpu-resourceclaim-template"
)

var (
podResource = metav1.GroupVersionResource{Version: "v1", Resource: "pods"}
)

func applyGPUMutation(req *admissionv1.AdmissionRequest) ([]patchOperation, error) {
// Only mutate Pod CREATE
// Swati: may be add UPDATE
if req.Resource != podResource || req.Operation != admissionv1.Create {
klog.Infof("skip mutation for %v/%v", req.Resource, req.Operation)
return nil, nil
}

var pod corev1.Pod
if _, _, err := universalDeserializer.Decode(req.Object.Raw, nil, &pod); err != nil {
klog.Errorf("failed to decode Pod: %v", err)
return nil, fmt.Errorf("could not deserialize pod: %w", err)
}

key := escapeJSONPointer(gpuResourceName)
var patches []patchOperation
var ctrGPUResourceClaims []string

// Iterate on all containers and check for "nvidia.com/gpu" limits
// using the logic described here for prefering limits over requests
// GPUs are only supposed to be specified in the limits section, meaning
// - can specify GPU limits without specifying requests. limit will be used as request value by default
// - can specify GPU in both limits and requests but they must be equal
// - cannot specify GPU requests without specifying limits
// refer: https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/#using-device-plugins
for ci, ctr := range pod.Spec.Containers {
ctrName := ctr.Name
limitCount, limitOk := ctr.Resources.Limits[gpuResourceName]

// skip if no GPUs in limits
if !limitOk || limitCount.Value() < 1 {
continue
}
gpuCount := limitCount.Value()

// check any GPUs in requests
// it must be equal to limits
if reqCount, reqOK := ctr.Resources.Requests[gpuResourceName]; reqOK {
if reqCount.Value() != gpuCount {
klog.Warningf("container[%q]: gpu request (%d) != limit (%d), skipping mutation", ctrName, reqCount.Value(), gpuCount)
continue
}
reqPatch := removeResourceRequest(ci, "requests", key)
patches = append(patches, reqPatch)
klog.Infof("removed container[%q].Resources.Requests: %v", ctrName, reqPatch)
}
limitPatch := removeResourceRequest(ci, "limits", key)
patches = append(patches, limitPatch)
klog.Infof("removed container[%q].Resources.Limits: %v", ctrName, limitPatch)

// ensure container-claims slice exists
// this is JSON way to first creating the field if it does not exist and append later with "-"
if len(ctr.Resources.Claims) == 0 {
createPatch := createClaimPatch(fmt.Sprintf("/spec/containers/%d/resources/claims", ci))
patches = append(patches, createPatch)
klog.Infof("created container[%q] empty claims array: %v", ctrName, createPatch)
}

// append one claim per GPU
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Updated code test:

$ kubectl logs gpu-mutating-webhook-7f676685c8-8w5sf -n nvidia-dra-driver-gpu
2025/05/02 19:52:34 Handling webhook request ...
I0502 19:52:34.675245       1 main.go:49] skip mutation for { v1 pods}/UPDATE
2025/05/02 19:52:34 Webhook request handled successfully
2025/05/02 19:53:05 Handling webhook request ...
I0502 19:53:05.135191       1 main.go:89] removed container["main"].Resources.Requests: {remove /spec/containers/0/resources/requests/nvidia.com~1gpu <nil>}
I0502 19:53:05.135219       1 main.go:93] removed container["main"].Resources.Limits: {remove /spec/containers/0/resources/limits/nvidia.com~1gpu <nil>}
I0502 19:53:05.135226       1 main.go:100] created container["main"] empty claims array: {add /spec/containers/0/resources/claims []}
I0502 19:53:05.135236       1 main.go:112] added to container["main"].Resources.Claims: {add /spec/containers/0/resources/claims/- map[name:nvidia-gpu-resourceclaim-0]}
I0502 19:53:05.135245       1 main.go:112] added to container["main"].Resources.Claims: {add /spec/containers/0/resources/claims/- map[name:nvidia-gpu-resourceclaim-1]}
I0502 19:53:05.135249       1 main.go:123] created pod["swati-gpu-pod"] empty claims array: {add /spec/resourceClaims []}
I0502 19:53:05.135256       1 main.go:136] added ResourceClaim "nvidia-gpu-resourceclaim-0" (template="nvidia-gpu-resourceclaim-template") to "swati-gpu-pod": {add /spec/resourceClaims/- map[name:nvidia-gpu-resourceclaim-0 resourceClaimTemplateName:nvidia-gpu-resourceclaim-template]}
I0502 19:53:05.135264       1 main.go:136] added ResourceClaim "nvidia-gpu-resourceclaim-1" (template="nvidia-gpu-resourceclaim-template") to "swati-gpu-pod": {add /spec/resourceClaims/- map[name:nvidia-gpu-resourceclaim-1 resourceClaimTemplateName:nvidia-gpu-resourceclaim-template]}
2025/05/02 19:53:05 Webhook request handled successfully

$ kubectl get resourceclaim 
NAME                                             STATE                AGE
swati-gpu-pod-nvidia-gpu-resourceclaim-0-x6dln   allocated,reserved   3m10s
swati-gpu-pod-nvidia-gpu-resourceclaim-1-hkbgk   allocated,reserved   3m10s

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could we construct unit tests that exercise the same logic?

for i := int64(0); i < gpuCount; i++ {
claimName := fmt.Sprintf("%s-%d", gpuClaimName, i)
ctrGPUResourceClaims = append(ctrGPUResourceClaims, claimName)
appendPatch := appendClaimPatch(
fmt.Sprintf("/spec/containers/%d/resources/claims", ci),
map[string]string{"name": claimName},
)
patches = append(patches, appendPatch)
klog.Infof("added to container[%q].Resources.Claims: %v", ctrName, appendPatch)
}
}

// Add claims pod-level
podName := pod.Name
if len(ctrGPUResourceClaims) > 0 {
// ensure pod-claims slice exists
if len(pod.Spec.ResourceClaims) == 0 {
createPatch := createClaimPatch("/spec/resourceClaims")
patches = append(patches, createPatch)
klog.Infof("created pod[%q] empty claims array: %v", podName, createPatch)
}

// append each container GPU claim at pod-level
for _, name := range ctrGPUResourceClaims {
appendPatch := appendClaimPatch(
"/spec/resourceClaims",
map[string]string{
"name": name,
"resourceClaimTemplateName": gpuTemplateName,
},
)
patches = append(patches, appendPatch)
klog.Infof("added ResourceClaim %q (template=%q) to %q: %v", name, gpuTemplateName, podName, appendPatch)
}
}

return patches, nil
}

// escapeJSONPointer replace "/" with "~1"
// refer: https://github.com/json-patch/json-patch-tests/issues/42
// needed for "nvidia.com/gpu". otherwise JSON will treat "/" as a path delimiter and treat "gpu" as new field
func escapeJSONPointer(s string) string {
return strings.ReplaceAll(s, "/", "~1")
}

// removeResourceRequest removes either .resources.requests or .resources.limits
func removeResourceRequest(ci int, field, key string) patchOperation {
return patchOperation{
Op: "remove",
Path: fmt.Sprintf("/spec/containers/%d/resources/%s/%s", ci, field, key),
}
}

// createClaimPatch creates an empty slice at the given path
func createClaimPatch(path string) patchOperation {
return patchOperation{
Op: "add",
Path: path,
Value: []map[string]string{},
}
}

// appendClaimPatch appends to the slice at path
// "-" is JSON way to inserting at the end of the array when no index is specified.
// refer: https://datatracker.ietf.org/doc/html/rfc6902
func appendClaimPatch(path string, entry map[string]string) patchOperation {
return patchOperation{
Op: "add",
Path: path + "/-",
Value: entry,
}
}

func main() {
certPath := filepath.Join(tlsDir, tlsCertFile)
keyPath := filepath.Join(tlsDir, tlsKeyFile)

mux := http.NewServeMux()
mux.Handle("/mutate", admitFuncHandler(applyGPUMutation))

server := &http.Server{
Addr: ":8443",
Handler: mux,
}

if err := server.ListenAndServeTLS(certPath, keyPath); err != nil {
// Swati: need better error handling here
log.Fatalf("Failed to start server: %v", err)
}
klog.Infof("Started gpu-mutating-webhook server at %s", server.Addr)
}
1 change: 1 addition & 0 deletions deployments/container/Dockerfile
Original file line number Diff line number Diff line change
Expand Up @@ -67,4 +67,5 @@ RUN mkdir /licenses && mv /NGC-DL-CONTAINER-LICENSE /licenses/NGC-DL-CONTAINER-L
COPY --from=build /artifacts/compute-domain-controller /usr/bin/compute-domain-controller
COPY --from=build /artifacts/compute-domain-kubelet-plugin /usr/bin/compute-domain-kubelet-plugin
COPY --from=build /artifacts/gpu-kubelet-plugin /usr/bin/gpu-kubelet-plugin
COPY --from=build /artifacts/gpu-mutating-webhook /usr/bin/gpu-mutating-webhook
COPY --from=build /build/templates /templates
Loading
Loading