Merge pull request #198 from klueska/explicit-envvar-for-mask

Add explicit envvar to control if we mask /proc/driver/nvidia/params

Author: klueska

demo/clusters/kind/install-dra-driver.sh

@@ -29,6 +29,7 @@ helm upgrade -i --create-namespace --namespace nvidia nvidia-dra-driver ${PROJEC
     --set deviceClasses="{${deviceClasses}}" \
     ${NVIDIA_CTK_PATH:+--set nvidiaCtkPath=${NVIDIA_CTK_PATH}} \
     ${NVIDIA_DRIVER_ROOT:+--set nvidiaDriverRoot=${NVIDIA_DRIVER_ROOT}} \
+    ${MASK_NVIDIA_DRIVER_PARAMS:+--set maskNvidiaDriverParams=${MASK_NVIDIA_DRIVER_PARAMS}} \
     --wait
 
 set +x

demo/clusters/nvkind/install-dra-driver.sh

@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+
+# 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.
+
+# A reference to the current directory where this script is located
+CURRENT_DIR="$(cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)"
+
+: ${MASK_NVIDIA_DRIVER_PARAMS:="true"}
+export MASK_NVIDIA_DRIVER_PARAMS
+exec ${CURRENT_DIR}/../kind/install-dra-driver.sh

demo/clusters/nvkind/install-dra-driver.sh

@@ -56,14 +56,11 @@ spec:
         command: ["bash", "-c"]
         args:
         - |-
-          # TODO: Masking of the params file is done below to allow nvkind to
-          # selectively exclude certain GPUs from being visible to the driver.
-          # At present, this is only feasible with a host-mounted driver where
-          # /dev in this container already has GPU devices present (as brought
-          # in via the --privileged flag from docker/podman when using nvkind).
-          # In the future we should revisit this to find a more robust method
-          # of supporting this.
-          if [ "${NVIDIA_DRIVER_ROOT}" = "/" ]; then
+          # Conditionally mask the params file to prevent this container from
+          # recreating any missing GPU device nodes. This is necessary, for
+          # example, when running under nvkind to limit the set GPUs governed
+          # by the plugin even though it has cgroup access to all of them.
+          if [ "${MASK_NVIDIA_DRIVER_PARAMS}" = "true" ]; then
             cp /proc/driver/nvidia/params root/gpu-params
             sed -i 's/^ModifyDeviceFiles: 1$/ModifyDeviceFiles: 0/' root/gpu-params
             mount --bind root/gpu-params /proc/driver/nvidia/params
@@ -72,6 +69,8 @@ spec:
         resources:
           {{- toYaml .Values.kubeletPlugin.containers.plugin.resources | nindent 10 }}
         env:
+        - name: MASK_NVIDIA_DRIVER_PARAMS
+          value: "{{ .Values.maskNvidiaDriverParams }}"
         - name: NVIDIA_CTK_PATH
           value: "{{ .Values.nvidiaCtkPath }}"
         - name: NVIDIA_DRIVER_ROOT

deployments/helm/k8s-dra-driver/templates/kubeletplugin.yaml

@@ -36,6 +36,17 @@ allowDefaultNamespace: false
 
 deviceClasses: ["gpu", "mig", "imex"]
 
+# Masking of the params file is typically done to allow nvkind to
+# selectively exclude certain GPUs from being visible to the
+# underlying GPU driver. Unfortunately, kind doesn't let you choose
+# which device nodes to inject into each worker node (they all come in
+# via the --priviliged flag passed to docker/podman). Because of
+# this, all workers see all GPUs by default. By masking the params
+# file we can prevent a container from recreating any missing GPU
+# device nodes and limit its view to only those device nodes that
+# nvkind decided to allow in.
+maskNvidiaDriverParams: false
+
 imagePullSecrets: []
 image:
   repository: nvcr.io/nvidia/cloud-native/k8s-dra-driver