[DCV] Fix Xdcv segfault on RHEL9 and Rocky9 caused by DCV attempting GL initialization when GPU acceleration is not supported.

CHANGELOG.md

@@ -39,6 +39,7 @@ This file is used to list changes made in each version of the AWS ParallelCluste
 - Upgrade amazon-efs-utils to version 2.4.0 (from v2.1.0) for Amazon Linux AMI's.
 
 **BUG FIXES**
+- Fix Xdcv segfault on RHEL9 and Rocky9 caused by DCV attempting GL initialization when GPU acceleration is not supported.
 - Fix a failure when creating a cluster with GPU instances and with DCV enabled but without internet access.
 - Fix build-image failure during ubuntu-desktop installation on a Ubuntu parent image with outdated OS packages.
 - Fix the CloudWatch agent configuration to ensure proper parsing of timestamps across all log files.

cookbooks/aws-parallelcluster-platform/resources/dcv/partial/_dcv_common.rb

@@ -188,7 +188,8 @@ def optionally_disable_rnd
 
 action :configure do
   if dcv_supported? && (node['cluster']['node_type'] == "HeadNode" || node['cluster']['node_type'] == "LoginNode")
-    if dcv_gpu_accel_supported?
+    gpu_accel = dcv_gpu_accel_supported?
+    if gpu_accel
       # Enable graphic acceleration in dcv conf file for graphic instances.
       allow_gpu_acceleration
     else
@@ -234,6 +235,7 @@ def optionally_disable_rnd
       owner 'root'
       group 'root'
       mode '0755'
+      variables(gpu_accel_supported: gpu_accel)
     end
 
     # Create directory for the external authenticator to store access file created by the users

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

@@ -931,6 +931,12 @@ def self.nothing(chef_run)
         it 'starts Amazon DCV server' do
           is_expected.to enable_service('dcvserver').with_action(%i(enable start))
         end
+
+        it 'passes gpu_accel_supported=true to dcv.conf template' do
+          is_expected.to create_template('/etc/dcv/dcv.conf').with(
+            variables: { gpu_accel_supported: true }
+          )
+        end
       end
 
       context "when g2 instance" do
@@ -1001,6 +1007,12 @@ def self.nothing(chef_run)
             .with_code(/systemctl set-default graphical.target/)
             .with_code(/systemctl isolate graphical.target &/)
         end
+
+        it 'passes gpu_accel_supported=false to dcv.conf template' do
+          is_expected.to create_template('/etc/dcv/dcv.conf').with(
+            variables: { gpu_accel_supported: false }
+          )
+        end
       end
     end
   end

cookbooks/aws-parallelcluster-platform/templates/dcv/dcv.conf.erb

@@ -41,7 +41,13 @@
 # 'dcv-gl' feature (a specific license will be required).
 # Allowed values: 'always-on', 'always-off', 'default-on', 'default-off'.
 # If not specified, the default value is 'default-on'.
+<% if @gpu_accel_supported %>
 #enable-gl-in-virtual-sessions = "default-on"
+<% else %>
+# Explicitly disable dcv-gl on instances where GPU acceleration is not supported
+# (e.g. g5g) to prevent Xdcv segfaults during GLX initialization.
+enable-gl-in-virtual-sessions = "always-off"
+<% end %>
 
 # Property "virtual-session-source-profile" specifies whether the shell that
 # runs the session starter script should source the user profile.
@@ -99,7 +105,10 @@ virtual-session-source-profile = true
 # second that are sent to the client. A higher value consumes more bandwidth
 # and resources. By default it is set to 25. Set to 0 for no limit
 #target-fps = 30
-<% unless node['cluster']['dcv']['is_graphic_instance'] %>
+<% unless @gpu_accel_supported %>
+# On instances that do not support GPU acceleration, DCV must use
+# software-only encoders because there is no GPU that can be used
+# for hardware encoding.
 display-encoders=['turbojpeg', 'lz4', 'ffmpeg']
 <% end %>
 

cookbooks/aws-parallelcluster-platform/test/controls/dcv_spec.rb

@@ -238,6 +238,11 @@
     it { should be_owned_by 'root' }
     it { should be_grouped_into 'root' }
     it { should be_mode 0755 }
+
+    unless instance.graphic? && instance.nvidia_installed? && instance.dcv_gpu_accel_supported?
+      its('content') { should match /enable-gl-in-virtual-sessions\s*=\s*"always-off"/ }
+      its('content') { should match /display-encoders=\['turbojpeg', 'lz4', 'ffmpeg'\]/ }
+    end
   end
 
   describe directory('/var/spool/parallelcluster/pcluster_dcv_authenticator') do