gpu: nvidia install drops VFIO autoload; GPU flows self-repair stale vfio.conf

The nvidia driver installer now writes only `nvidia` and `nvidia_uvm` into `/etc/modules-load.d/nvidia-vfio.conf`, leaving the vfio-pci modules for `switch_gpu_mode.sh` to add and remove when the user explicitly transitions between LXC and VM mode. A host with residual `/etc/modprobe.d/vfio.conf` entries from an earlier VM passthrough setup no longer has vfio-pci racing nvidia at boot and capturing the GPU by ID.

All three GPU-facing entry points also detect a stale `vfio.conf` entry whose vid:did belongs to a GPU currently bound to a non-vfio driver, and offer to clean it in place. `nvidia_installer.sh` runs the check as a pre-flight before install. `switch_gpu_mode.sh` presents an interactive dialog and rebuilds initramfs. `switch_gpu_mode_direct.sh` performs the same cleanup transparently for the Monitor web variant. Once cleaned, "Add GPU to LXC", the driver install and Switch Mode all proceed without any manual intervention.

The two supporting helpers `_pci_driver_of` and `_clean_vfio_conf_ids` previously local to `add_gpu_lxc.sh` move to `scripts/global/pci_passthrough_helpers.sh`, so every GPU-related script resolves them from a single source without duplication.
This commit is contained in:
MacRimi
2026-08-17 12:05:40 +02:00
parent 11e030b9de
commit b5476c75fa
5 changed files with 199 additions and 62 deletions
+51
View File
@@ -563,8 +563,59 @@ _selected_gpu_current_mode() {
echo "$mode"
}
check_stale_vfio_config_switch_mode() {
local vfio_conf="/etc/modprobe.d/vfio.conf"
[[ ! -f "$vfio_conf" ]] && return 0
local ids_line ids_part
ids_line=$(grep "^options vfio-pci ids=" "$vfio_conf" 2>/dev/null | head -1)
[[ -z "$ids_line" ]] && return 0
ids_part=$(echo "$ids_line" | grep -oE 'ids=[^[:space:]]+' | sed 's/ids=//')
[[ -z "$ids_part" ]] && return 0
local -a legacy_ids=()
local legacy_list=""
local idx viddid drv pci name
for idx in "${SELECTED_GPU_IDX[@]}"; do
viddid="${ALL_GPU_VIDDID[$idx]}"
drv="${ALL_GPU_DRIVERS[$idx]}"
pci="${ALL_GPU_PCIS[$idx]}"
name="${ALL_GPU_NAMES[$idx]}"
[[ -z "$viddid" ]] && continue
[[ "$drv" == "vfio-pci" ]] && continue
if echo ",${ids_part}," | grep -q ",${viddid},"; then
legacy_ids+=("$viddid")
legacy_list+="${name} (${pci}) [${viddid}]\n"
fi
done
[[ ${#legacy_ids[@]} -eq 0 ]] && return 0
local msg
msg="\n$(translate 'The following selected GPU(s) still have a VFIO passthrough entry in') /etc/modprobe.d/vfio.conf:\n\n"
msg+="${legacy_list}\n"
msg+="$(translate 'The active kernel driver is not vfio-pci, but the entry will rebind the GPU to vfio-pci on the next reboot.')\n\n"
msg+="\Z1\Zb$(translate 'Do you want to remove the stale entry from vfio.conf?')\Zn"
dialog --colors --backtitle "ProxMenux" \
--title "$(translate 'Stale VFIO Config Detected')" \
--yesno "$msg" 18 80 || return 0
if declare -F _clean_vfio_conf_ids >/dev/null 2>&1 \
&& _clean_vfio_conf_ids "${legacy_ids[@]}"; then
msg_info "$(translate 'Rebuilding initramfs after vfio.conf cleanup...')"
update-initramfs -u >/dev/null 2>&1 || true
dialog --backtitle "ProxMenux" \
--title "$(translate 'Stale VFIO Config Cleaned')" \
--msgbox "\n$(translate 'The vfio.conf entries have been removed and initramfs rebuilt.')\n\n$(translate 'A reboot is recommended before the GPU is guaranteed to stay on the native driver.')" \
12 78
exit 0
fi
}
select_target_mode() {
CURRENT_MODE=$(_selected_gpu_current_mode)
check_stale_vfio_config_switch_mode
if [[ "$CURRENT_MODE" == "mixed" ]]; then
local msg idx mode_label