non-slash driver root: mount parent

Signed-off-by: Dr. Jan-Philip Gehrcke <[email protected]>

Author: jgehrcke

deployments/helm/nvidia-dra-driver-gpu/templates/kubeletplugin.yaml

@@ -61,18 +61,25 @@ spec:
         # Use runc: explicit "void"; otherwise we inherit "all".
         - name: NVIDIA_VISIBLE_DEVICES
           value: void
-        # In case of the operator-provided driver, another container mounts the
-        # driver onto the host using `mountPropagation: Bidirectional`
-        # (out-of-band of the lifecycle of _this_ pod here). For us to see that
-        # mount, `mountPropagation: HostToContainer` is required (docs: "if any
-        # Pod with Bidirectional mount propagation to the same volume mounts
-        # anything there, the container with HostToContainer mount propagation
-        # will see it.").
         volumeMounts:
-        - name: driver-root
+      {{- if eq "/" .Values.nvidiaDriverRoot }}
+        volumeMounts:
+        - name: host-root
           mountPath: /driver-root
           readOnly: true
+      {{- else }}
+        volumeMounts:
+        - name: driver-root-parent
+          mountPath: /driver-root-parent
+          # In case of the operator-provided driver, another container mounts
+          # the driver onto the host using `mountPropagation: Bidirectional`
+          # (out-of-band of the lifecycle of _this_ pod here). For us to see
+          # that mount, `mountPropagation: HostToContainer` is required (docs:
+          # "if any Pod with Bidirectional mount propagation to the same volume
+          # mounts anything there, the container with HostToContainer mount
+          # propagation will see it.").
           mountPropagation: HostToContainer
+      {{- end }}
       containers:
       {{- if .Values.resources.computeDomains.enabled }}
       - name: compute-domains
@@ -197,6 +204,19 @@ spec:
       - name: cdi
         hostPath:
           path: /var/run/cdi
+      {{- if eq "/" .Values.nvidiaDriverRoot }}
+      - name: host-root
+        hostPath:
+          path: /
+      {{- else }}
+      - name: driver-root-parent
+        hostPath:
+          # First, remove trailing slash (if exists) and then remove last path
+          # element with sprig template function `dir`. Examples: /a/b/ -> /a,
+          # /a/b/c -> /a/b.
+          path: {{ dir (trimSuffix "/" .Values.nvidiaDriverRoot) }}
+          type: DirectoryOrCreate
+      {{- end }}
       - name: driver-root
         hostPath:
           path: {{ .Values.nvidiaDriverRoot }}

hack/kubelet-plugin-prestart.sh

@@ -11,6 +11,20 @@ if [ -z "$NVIDIA_DRIVER_ROOT" ]; then
     export NVIDIA_DRIVER_ROOT="/"
 fi
 
+
+if [ "${NVIDIA_DRIVER_ROOT}" != "/" ]; then
+    # Create in-container path /driver-root as a symlink. Expectation: link may
+    # be broken initially (e.g., if the GPU operator isn't deployed yet. The
+    # link heals the driver becomes mounted (e.g., once GPU operator provides
+    # the driver on the host at /run/nvidia/driver).
+
+    # Remove trailing slash (if existing) and get last elem: /a/b/c -> c,
+    # /a/b/c/ -> c
+    _driver_root_basename=$("${NVIDIA_DRIVER_ROOT%/}" | xargs basename)
+    echo "create symlink: /driver-root -> /driver-root-parent/${_driver_root_basename}"
+    ln -s "/driver-root-parent/${_driver_root_basename}" /driver-root
+fi
+
 emit_common_err () {
     printf '%b' \
         "Check failed. Has the NVIDIA GPU driver been set up? " \