Fix issues when performing OFED/CUDA/Lustre extra builds

ansible/roles/cuda/defaults/main.yml

@@ -1,6 +1,6 @@
 cuda_distro: "rhel{{ ansible_distribution_major_version }}"
 cuda_repo: "https://developer.download.nvidia.com/compute/cuda/repos/{{ cuda_distro }}/x86_64/cuda-{{ cuda_distro }}.repo"
-cuda_driver_stream: default
+cuda_driver_stream: 560-open
 cuda_package_version: 'latest'
 cuda_packages:
   - "cuda{{ ('-' + cuda_package_version) if cuda_package_version != 'latest' else '' }}"

ansible/roles/cuda/tasks/main.yml

@@ -25,7 +25,7 @@
   register: _cuda_driver_module_enabled
 
 - name: Enable nvidia driver module
-  ansible.builtin.command: "dnf module enable -y nvidia-driver:open-dkms"
+  ansible.builtin.command: "dnf module enable -y nvidia-driver:{{ cuda_driver_stream }}"
   register: _cuda_driver_module_enable
   when: "'No matching Modules to list' in _cuda_driver_module_enabled.stderr"
   changed_when: "'Nothing to do' not in _cuda_driver_module_enable.stdout"

ansible/roles/lustre/tasks/install.yml

@@ -42,29 +42,8 @@
     name: "{{ _lustre_find_rpms.files | map(attribute='path')}}"
     disable_gpg_check: yes
 
-- block:
-  - name: Remove lustre build prerequisites
-    # NB Only remove ones this role installed which weren't upgrades
-    ansible.builtin.dnf:
-      name: "{{ _new_pkgs }}"
-      state: absent
-    vars:
-      _installed_pkgs: |
-        {{
-          _lustre_dnf_build_packages.results |
-          select('match', 'Installed:') |
-          map('regex_replace', '^Installed: (.+?)-[0-9].*$', '\1')
-        }}
-      _removed_pkgs: |
-        {{
-          _lustre_dnf_build_packages.results |
-          select('match', 'Removed:') |
-          map('regex_replace', '^Removed: (.+?)-[0-9].*$', '\1')
-        }}
-      _new_pkgs: "{{ _installed_pkgs | difference(_removed_pkgs) }}"
-
-  - name: Delete lustre build dir
-    file:
+- name: Delete lustre build dir
+  file:
       path: "{{ lustre_build_dir }}"
       state: absent
   when: lustre_build_cleanup | bool

ansible/roles/ofed/tasks/install.yml

@@ -1,3 +1,13 @@
+
+- name: Install latest kernel packages
+  ansible.builtin.dnf:
+    name:
+    - kernel
+    - kernel-core
+    - kernel-tools
+    - kernel-tools-libs
+    state: latest
+
 - name: Get installed kernels
   command: dnf list --installed kernel
   register: _ofed_dnf_kernels
@@ -8,16 +18,20 @@
   register: _ofed_loaded_kernel
   changed_when: false
 
-- name: Check current kernel is newest installed
-  assert:
-    that: _ofed_kernel_current == _ofed_dnf_kernels_newest
-    fail_msg: "Kernel {{ _ofed_loaded_kernel.stdout }} is loaded but newer {{ _ofed_dnf_kernels_newest }} is installed: consider rebooting?"
+- name: Reboot into new kernel if not on latest
+  ansible.builtin.reboot:
   vars:
     _ofed_kernel_current: >-
       {{ _ofed_loaded_kernel.stdout | regex_replace('\.(?:.(?!\.))+$', '') | regex_replace('\.(?:.(?!\.))+$', '') }}
     _ofed_dnf_kernels_newest: >-
       {{ _ofed_dnf_kernels.stdout_lines[1:] | map('split') | map(attribute=1) | map('regex_replace', '\.(?:.(?!\.))+$', '') | community.general.version_sort | last }}
     # dnf line format e.g. "kernel.x86_64  4.18.0-513.18.1.el8_9   @baseos  "
+  when: _ofed_kernel_current != _ofed_dnf_kernels_newest
+
+- name: Get new running kernel
+  command: uname -r
+  register: _ofed_loaded_kernel
+  changed_when: false
 
 - name: Enable epel
   dnf: