From 730d10449b24276d623a9f246171a67fe4efd13f Mon Sep 17 00:00:00 2001
From: Piotr Mlocek <pmlocek@nvidia.com>
Date: Wed, 1 Apr 2026 12:05:12 -0700
Subject: [PATCH] fix(gpu): gate Tegra-specific runtime adjustments

---
 crates/openshell-bootstrap/src/docker.rs      |  2 +-
 crates/openshell-sandbox/src/process.rs       | 74 +++++++++++++++----
 .../nvidia-device-plugin-helmchart.yaml       |  3 -
 3 files changed, 62 insertions(+), 17 deletions(-)

diff --git a/crates/openshell-bootstrap/src/docker.rs b/crates/openshell-bootstrap/src/docker.rs
index 403a680f7..4c08b35b1 100644
--- a/crates/openshell-bootstrap/src/docker.rs
+++ b/crates/openshell-bootstrap/src/docker.rs
@@ -603,7 +603,7 @@ pub async fn ensure_container(
     // nvidia runtime or CDI spec generation running inside k3s (for sandbox
     // pods) can apply the same config.
     const HOST_FILES_DIR: &str = "/etc/nvidia-container-runtime/host-files-for-container.d";
-    if std::path::Path::new(HOST_FILES_DIR).is_dir() {
+    if !device_ids.is_empty() && std::path::Path::new(HOST_FILES_DIR).is_dir() {
         let mut binds = host_config.binds.take().unwrap_or_default();
         binds.push(format!("{HOST_FILES_DIR}:{HOST_FILES_DIR}:ro"));
         host_config.binds = Some(binds);
diff --git a/crates/openshell-sandbox/src/process.rs b/crates/openshell-sandbox/src/process.rs
index 9726e4b75..8de5a4fcd 100644
--- a/crates/openshell-sandbox/src/process.rs
+++ b/crates/openshell-sandbox/src/process.rs
@@ -34,6 +34,25 @@ fn scrub_sensitive_env(cmd: &mut Command) {
     cmd.env_remove(SSH_HANDSHAKE_SECRET_ENV);
 }
 
+#[cfg(not(any(
+    target_os = "macos",
+    target_os = "ios",
+    target_os = "haiku",
+    target_os = "redox"
+)))]
+fn has_gpu_devices() -> bool {
+    const GPU_DEVICE_PATHS: &[&str] = &[
+        "/dev/nvidiactl",
+        "/dev/nvidia-uvm",
+        "/dev/nvidia-uvm-tools",
+        "/dev/nvidia-modeset",
+        "/dev/nvmap",
+    ];
+    GPU_DEVICE_PATHS
+        .iter()
+        .any(|path| std::path::Path::new(path).exists())
+}
+
 /// Handle to a running process.
 pub struct ProcessHandle {
     child: Child,
@@ -414,26 +433,33 @@ pub fn drop_privileges(policy: &SandboxPolicy) -> Result<()> {
             target_os = "redox"
         )))]
         {
+            let preserve_gpu_groups = has_gpu_devices();
             // Snapshot the container-level supplemental GIDs (e.g. injected by
             // CDI for GPU device access) before initgroups replaces them.
             // Exclude GID 0 (root) to avoid inadvertent privilege retention.
             let root_gid = nix::unistd::Gid::from_raw(0);
-            let container_gids: Vec<nix::unistd::Gid> = nix::unistd::getgroups()
-                .unwrap_or_default()
-                .into_iter()
-                .filter(|&g| g != root_gid)
-                .collect();
+            let container_gids: Vec<nix::unistd::Gid> = if preserve_gpu_groups {
+                nix::unistd::getgroups()
+                    .unwrap_or_default()
+                    .into_iter()
+                    .filter(|&g| g != root_gid)
+                    .collect()
+            } else {
+                Vec::new()
+            };
             nix::unistd::initgroups(user_cstr.as_c_str(), group.gid).into_diagnostic()?;
-            // Merge back any CDI-injected GIDs that initgroups dropped so that
-            // exec'd processes retain access to GPU devices (e.g. /dev/nvmap on
-            // Tegra requires the video GID).
-            let mut merged: Vec<nix::unistd::Gid> = nix::unistd::getgroups().unwrap_or_default();
-            for gid in container_gids {
-                if !merged.contains(&gid) {
-                    merged.push(gid);
+            if preserve_gpu_groups {
+                // Merge back CDI-injected GIDs that initgroups dropped so that
+                // exec'd processes retain access to GPU devices.
+                let mut merged: Vec<nix::unistd::Gid> =
+                    nix::unistd::getgroups().unwrap_or_default();
+                for gid in container_gids {
+                    if !merged.contains(&gid) {
+                        merged.push(gid);
+                    }
                 }
+                nix::unistd::setgroups(&merged).into_diagnostic()?;
             }
-            nix::unistd::setgroups(&merged).into_diagnostic()?;
         }
     }
 
@@ -629,6 +655,28 @@ mod tests {
         );
     }
 
+    #[cfg(not(any(
+        target_os = "macos",
+        target_os = "ios",
+        target_os = "haiku",
+        target_os = "redox"
+    )))]
+    #[test]
+    fn has_gpu_devices_detects_present_nodes() {
+        assert_eq!(
+            has_gpu_devices(),
+            [
+                "/dev/nvidiactl",
+                "/dev/nvidia-uvm",
+                "/dev/nvidia-uvm-tools",
+                "/dev/nvidia-modeset",
+                "/dev/nvmap",
+            ]
+            .iter()
+            .any(|path| std::path::Path::new(path).exists())
+        );
+    }
+
     #[tokio::test]
     async fn scrub_sensitive_env_removes_ssh_handshake_secret() {
         let mut cmd = Command::new("/usr/bin/env");
diff --git a/deploy/kube/gpu-manifests/nvidia-device-plugin-helmchart.yaml b/deploy/kube/gpu-manifests/nvidia-device-plugin-helmchart.yaml
index 737efdb54..1cb0ca70a 100644
--- a/deploy/kube/gpu-manifests/nvidia-device-plugin-helmchart.yaml
+++ b/deploy/kube/gpu-manifests/nvidia-device-plugin-helmchart.yaml
@@ -31,9 +31,6 @@ spec:
   targetNamespace: nvidia-device-plugin
   createNamespace: true
   valuesContent: |-
-    image:
-      repository: ghcr.io/nvidia/k8s-device-plugin
-      tag: "2ab68c16"
     runtimeClassName: nvidia
     deviceListStrategy: cdi-cri
     deviceIDStrategy: index