[BuildImage] Load kernel module drm_client_lib before the installation of NVIDIA driver, if the module is available on the kernel.

Starting kernel `5.14.0-611`, some DRM symbols required by the NVIDIA driver are exported by new client modules.

Author: gmarciani

CHANGELOG.md

@@ -7,8 +7,9 @@ This file is used to list changes made in each version of the AWS ParallelCluste
 ------
 
 **CHANGES**
-1. Add chef attribute `cluster/in_place_update_on_fleet_enabled` to disable in-place updates on compute and login nodes
-   and achieve better performance at scale.
+- Add chef attribute `cluster/in_place_update_on_fleet_enabled` to disable in-place updates on compute and login nodes
+   and achieve better performance at scale. 
+- Load kernel module `drm_client_lib` before installation of NVIDIA driver, if available on the kernel.
 
 
 3.14.0

cookbooks/aws-parallelcluster-platform/resources/nvidia_driver/partial/_nvidia_driver_common.rb

@@ -72,6 +72,13 @@
     end
   end
 
+  # Load kernel modules in best effort
+  kernel_modules_to_load.each do |kernel_module|
+    execute "Load kernel module if exposed by the kernel: #{kernel_module}" do
+      command "if modinfo #{kernel_module}; then modprobe #{kernel_module}; fi"
+    end
+  end
+
   # Install driver
   bash 'nvidia.run advanced' do
     user 'root'
@@ -126,3 +133,7 @@ def nvidia_kernel_module
     "kernel-open"
   end
 end
+
+def kernel_modules_to_load
+  %w(drm_client_lib)
+end

cookbooks/aws-parallelcluster-platform/spec/unit/resources/nvidia_driver_spec.rb

@@ -140,11 +140,27 @@ def self.setup(chef_run, nvidia_driver_version: nil)
   end
 end
 
+describe 'nvidia_driver:kernel_modules_to_load' do
+  cached(:chef_run) do
+    ChefSpec::SoloRunner.new(step_into: ['nvidia_driver'])
+  end
+
+  cached(:resource) do
+    ConvergeNvidiaDriver.setup(chef_run)
+    chef_run.find_resource('nvidia_driver', 'setup')
+  end
+
+  it 'returns expected kernel modules' do
+    expect(resource.kernel_modules_to_load).to eq(%w(drm_client_lib))
+  end
+end
+
 describe 'nvidia_driver:setup' do
   for_all_oses do |platform, version|
     cached(:nvidia_arch) { 'nvidia_arch' }
     cached(:nvidia_kernel_module) { 'nvidia_kernel_module' }
     cached(:nvidia_driver_version) { 'nvidia_driver_version' }
+    cached(:kernel_modules_to_load) { %w(module1 module2) }
     cached(:nvidia_driver_url) { "https://us.download.nvidia.com/tesla/#{nvidia_driver_version}/NVIDIA-Linux-#{nvidia_arch}-#{nvidia_driver_version}.run" }
 
     context "on #{platform}#{version} when nvidia_driver not enabled" do
@@ -176,6 +192,7 @@ def self.setup(chef_run, nvidia_driver_version: nil)
             allow(res).to receive(:nvidia_arch).and_return(nvidia_arch)
             allow(res).to receive(:nvidia_kernel_module).and_return(kernel_module)
             allow(res).to receive(:gcc_major_version_used_by_kernel).and_return(kernel_compiler_version)
+            allow(res).to receive(:kernel_modules_to_load).and_return(kernel_modules_to_load)
           end
 
           stub_command("lsinitramfs /boot/initrd.img-$(uname -r) | grep nouveau").and_return(true)
@@ -220,6 +237,14 @@ def self.setup(chef_run, nvidia_driver_version: nil)
           )
         end
 
+        it 'loads kernel modules in they are exposed by the kernel' do
+          kernel_modules_to_load.each do |kernel_module|
+            is_expected.to run_execute("Load kernel module if exposed by the kernel: #{kernel_module}").with(
+              command: "if modinfo #{kernel_module}; then modprobe #{kernel_module}; fi"
+            )
+          end
+        end
+
         if platform == 'amazon'
           compiler_version = version == '2023' ? 'gcc' : 'gcc10'
           compiler_path = version == '2023' ? 'CC=/usr/bin/gcc' : 'CC=/usr/bin/gcc10-gcc'