nvidia / dgx-selinux

DGX RHEL SELinux Policies

dgx-selinux

Introduction

This project contains the SELinux policies used in DGX products.

Building and debugging

The policy module builds require that the following packages are installed:

# RHEL7
sudo yum install -y findutils \
                    selinux-policy \
                    selinux-policy-devel \
                    selinux-policy-base \
                    libselinux-utils \
                    policycoreutils \
                    policycoreutils-python

# RHEL8/RHEL9
sudo dnf install -y findutils \
                    selinux-policy \
                    selinux-policy-devel \
                    selinux-policy-base \
                    libselinux-utils \
                    policycoreutils \
                    policycoreutils-python3

While this is not required, it may be useful for debugging or just evaluating the various policies on the system.

sudo yum install -y setools-console

setools-console provides the sesearch utility, which is very useful for querying installed modules and such:

sesearch --help
sesearch --allow | grep nvidia

Pre-compiled policy modules are also included in this project under the bin/ directory. Users can install these without having to compile the policy themselves if they so desired:

sudo semodule -i nvidia-container.pp

Information

The src/<policy-module-name> directories contain the source, Makefiles, and READMEs for each policy module. Definitions for the macros used in this project can be found in /usr/share/selinux/devel/include.

Disclaimer

These SELinux policies have only been verified on DGX-1 and DGX-2 on RHEL 7.5 and 7.6. They are meant to be used as a reference for RHEL-based non-DGX platforms, not verbatim.

dgx-selinux's Issues

nvidia-container policy seems to prevent connecting via SSH

Hi, I'm very happy that this SELinux policy exists! However, I seem to have found one problem:

My goal is to start a container that can access NVIDIA GPUs and that one can connect to via SSH:

1. I start my container via:

   $ podman run --rm -it -p 10022:22 \
       --security-opt label=type:nvidia_container_t \
       nvcr.io/nvidia/cuda:11.2.2-base-centos8

2. I can verify that GPUs are accessible, i.e., nvidia-smi returns the expected output.

3. I now want to set up and run an SSH server inside the container:

   [root@176645926b31 /]# yum -y install openssh-server
   [root@176645926b31 /]# ssh-keygen -A
   [root@176645926b31 /]# mkdir -p /root/.ssh
   [root@176645926b31 /]# echo "YOUR_SSH_PUBLIC_KEY" > /root/.ssh/authorized_keys
   [root@176645926b31 /]# /usr/sbin/sshd -D

4. And finally connect to it from outside. However, here, I receive the following error:

   $ ssh root@localhost -p 10022
   Connection closed by ::1 port 10022

If I repeat the above steps but without the --security-opt label=type:nvidia_container_t option, I can connect via SSH as expected (i.e., don't receive the error in the last step). Naturally, the nvidia-smi does not work in this case because it is blocked by SELinux as the nvidia-container policy is not loaded.

I am also seeing the same behavior with the nvcr.io/nvidia/cuda:11.2.2-base-centos7 image or the --cap-add ALL flag.

What is preventing the SSH connection here? Can the policy be changed to support this behavior? Or is there some option I can pass when starting the container to allow this? Thank you!

Prebuild policy, without the need to install the SELinux toolchain

Can we include a prebuild policy (the .pp file) as a release/download, so others can use the policy right away without installing SELinux toolchain?

RHEL 9 support

Given these selinux policies have been useful in RHEL7 with enabling rootless containers even on non-DGX systems, it would be useful if this project added official support for such policies in RHEL9