feat: NVIDIA modules in the initramfs only while every kernel has them, decided again by a pacman hook at every kernel change

CHANGELOG.md

@@ -11,6 +11,7 @@ one per merged pull request.
 - **test: tests/nvidia-test.sh runs the NVIDIA fix against a fake RTX 5080 beside an iGPU, a stub modinfo and temp copies of the Limine, systemd-boot and GRUB files (no root, VM or hardware); NVIDIA_DRM_DIR points detection at a fake sysfs tree**
 - **fix: NVIDIA fix on Limine edits the KERNEL_CMDLINE[default] line (an appended += line ended up as literal text on the kernel command line, found in a VM); the modules are only loaded early when every installed kernel has them, and a Limine boot entry without the parameters is an error**
 - **test: tests/nvidia-hardware-test.sh for a real NVIDIA PC: check (read-only state of the detection, the live nvidia_drm modeset/fbdev, versions), apply, and visual (was the gaming mode picture clean), with a report file to paste**
+- **fix: the NVIDIA modules are loaded early only while every installed kernel has them, decided again at every kernel or driver change by a pacman hook (steamify-nvidia-initramfs): a later kernel without the modules (another kernel, a failed DKMS build, a downgrade) no longer makes its initramfs and Limine boot entry fail**
 
 ## 2.9.6 - 2026-09-30
 

TECHNICAL.md

@@ -216,13 +216,18 @@ its driver installed (`lib/nvidia.sh`), the SteamOS conversion adds
 systemd-boot or GRUB; the file is backed up first) and
 `/etc/mkinitcpio.conf.d/90-steamify-nvidia.conf` (`MODULES+=(nvidia
 nvidia_modeset nvidia_uvm nvidia_drm)`), then rebuilds the initramfs and boot
-entries. The early-load drop-in is only written when every installed kernel
+entries. The early-load drop-in is only there while every installed kernel has the
-has the NVIDIA modules (mkinitcpio fails on a missing module, and
+NVIDIA modules (mkinitcpio fails on a missing module, and `limine-mkinitcpio`
-`limine-mkinitcpio` then skips that kernel's boot entry, parameters included).
+then skips that kernel's boot entry, parameters included). That is decided again
+at every kernel or driver change by a pacman hook
+(`/etc/pacman.d/hooks/85-steamify-nvidia-initramfs.hook`, after DKMS builds,
+before the initramfs is built) running `/usr/local/libexec/steamify-nvidia-initramfs`:
+another kernel without the modules, a failed DKMS build or a downgrade removes
+the drop-in, and it comes back once every kernel has them again.
 On Limine the parameters go into the existing `KERNEL_CMDLINE[default]="..."`
 line: an extra appended `+=` line is pasted into the command line as text.
 Only what's missing is changed; it applies after a reboot. Turning
-the conversion off removes both. `tests/nvidia-test.sh` tests it against a fake GPU
+the conversion off removes both, the hook and its script. `tests/nvidia-test.sh` tests it against a fake GPU
 and fake boot loader files; `tests/nvidia-hardware-test.sh check|apply|visual` is for a
 real NVIDIA PC (read-only check, the real apply, and a prompt for whether the picture
 was clean), and writes `~/steamify-nvidia-report.txt`.

lib/nvidia.sh

@@ -9,6 +9,9 @@
 
 NVIDIA_PARAMS="nvidia-drm.modeset=1 nvidia-drm.fbdev=1"
 NVIDIA_INITRAMFS_CONF=/etc/mkinitcpio.conf.d/90-steamify-nvidia.conf
+# The script and pacman hook that keep the initramfs file right at every kernel change.
+NVIDIA_SCRIPT=${NVIDIA_SCRIPT:-/usr/local/libexec/steamify-nvidia-initramfs}
+NVIDIA_HOOK=${NVIDIA_HOOK:-/etc/pacman.d/hooks/85-steamify-nvidia-initramfs.hook}
 # The DRM devices in sysfs; the tests point this at a fake tree.
 NVIDIA_DRM_DIR=${NVIDIA_DRM_DIR:-/sys/class/drm}
 
@@ -46,16 +49,14 @@ nvidia_initramfs_ok() {
 }
 
 nvidia_modules_everywhere() {
-    # 0 when every installed kernel has the nvidia_drm module: mkinitcpio fails
+    # 0 when every installed kernel has the nvidia_drm module (the script's own check).
-    # on a MODULES entry a kernel lacks, and limine-mkinitcpio then skips that
+    patch_file steamify-nvidia-initramfs.sh | NVIDIA_MODULES_DIR="${NVIDIA_MODULES_DIR:-/usr/lib/modules}" bash -s check
-    # kernel's initramfs and boot entry, so the parameters never reach it.
+}
-    local k found=0
+
-    for k in "${NVIDIA_MODULES_DIR:-/usr/lib/modules}"/*/; do
+nvidia_hook_install() {
-        [[ -f "$k/pkgbase" || -d "$k/kernel" || -d "$k/updates" || -d "$k/extramodules" ]] || continue
+    patch_file steamify-nvidia-initramfs.sh | sudo install -Dm755 /dev/stdin "$NVIDIA_SCRIPT" &&
-        find "$k" -name 'nvidia-drm.ko*' -print -quit 2>/dev/null | grep -q . || return 1
+        patch_file steamify-nvidia-initramfs.hook | fill SCRIPT="$NVIDIA_SCRIPT" |
-        found=1
+            sudo install -Dm644 /dev/stdin "$NVIDIA_HOOK"
-    done
-    [[ $found == 1 ]]
 }
 
 nvidia_rebuild_boot() {
@@ -107,12 +108,17 @@ nvidia_enable() {
             changed=1
         fi
     fi
-    if ! nvidia_initramfs_ok && ! nvidia_modules_everywhere; then
+    nvidia_hook_install || { err "Installing the NVIDIA initramfs hook failed."; return 1; }
-        warn "Not every installed kernel has the NVIDIA modules: skipping loading them early (the kernel parameters still apply)."
+    local had=0; [[ -f "$NVIDIA_INITRAMFS_CONF" ]] && had=1
-    elif ! nvidia_initramfs_ok; then
+    sudo env NVIDIA_INITRAMFS_CONF="$NVIDIA_INITRAMFS_CONF" NVIDIA_MODULES_DIR="${NVIDIA_MODULES_DIR:-/usr/lib/modules}" "$NVIDIA_SCRIPT" sync
+    if [[ -f "$NVIDIA_INITRAMFS_CONF" && $had == 0 ]]; then
         info "Loading the NVIDIA modules early (initramfs)..."
-        printf 'MODULES+=(nvidia nvidia_modeset nvidia_uvm nvidia_drm)\n' |
+        changed=1
-            sudo install -Dm644 /dev/stdin "$NVIDIA_INITRAMFS_CONF" && changed=1
+    elif [[ ! -f "$NVIDIA_INITRAMFS_CONF" && $had == 1 ]]; then
+        warn "A kernel without the NVIDIA modules is installed: not loading them early any more (rebuilding the initramfs without them)."
+        changed=1
+    elif [[ ! -f "$NVIDIA_INITRAMFS_CONF" ]] && ! nvidia_initramfs_ok; then
+        warn "Not every installed kernel has the NVIDIA modules: not loading them early for now (the kernel parameters still apply; this is checked again at every kernel update)."
     fi
     if [[ $changed == 1 ]]; then
         nvidia_rebuild_boot || { err "Rebuilding the boot entries failed."; return 1; }
@@ -132,6 +138,7 @@ nvidia_disable() {
         changed=1
     fi
     [[ -f "$NVIDIA_INITRAMFS_CONF" ]] && { sudo rm -f "$NVIDIA_INITRAMFS_CONF"; changed=1; }
+    sudo rm -f "$NVIDIA_HOOK" "$NVIDIA_SCRIPT"
     [[ $changed == 1 ]] && nvidia_rebuild_boot
     return 0
 }

patches/README.md

@@ -14,4 +14,5 @@ read and reviewed on their own. `lib/*.sh` reads them with `patch_file
 | `steam-shortcuts.py` | `lib/steam-game.sh` (Add as non-Steam game) | Reads and writes Steam's binary `shortcuts.vdf` (find, add or update, remove) |
 | `steamify-notifier.py` | `lib/update-notifier.sh` (Update notifications) | The daily check: notification and tray icon when there's a new release |
 | `steamify-os-release.sh`, `steamify-os-release.hook` | `lib/login-manager.sh` (SteamOS conversion) | What Steam's System settings show besides the OS name (which stays CachyOS's): `VERSION_ID=steamos-X.Y`, `VERSION_CODENAME=steam-machine`, `VARIANT`/`VARIANT_ID` Steamify with its version, in `/etc/os-release`; the pacman hook runs it again after cachyos-hooks (`fill SCRIPT=... VERSION=... STEAMOS=...`) |
+| `steamify-nvidia-initramfs.sh`, `steamify-nvidia-initramfs.hook` | `lib/nvidia.sh` (SteamOS conversion, NVIDIA fix) | The pacman hook (before the initramfs is built, after DKMS) and its script: the NVIDIA modules go into the initramfs only while every installed kernel has them, decided again at every kernel or driver change (`fill SCRIPT=...`) |
 | `cros-ec-cec-single-port.patch` | `lib/cec.sh` (HDMI-CEC) | Makes Valve's Steam Machine CEC driver find amdgpu's HDMI port |

patches/steamify-nvidia-initramfs.hook

@@ -0,0 +1,17 @@
+# Installed by Steamify CachyOS (SteamOS conversion, NVIDIA fix): decides
+# again whether the NVIDIA modules go into the initramfs, after DKMS has built
+# its modules (71-dkms-install) and before the initramfs is built
+# (90-mkinitcpio-install); hooks run in file name order.
+[Trigger]
+Operation = Install
+Operation = Upgrade
+Operation = Remove
+Type = Path
+Target = usr/lib/modules/*/vmlinuz
+Target = usr/lib/modules/*/extramodules/nvidia*
+Target = usr/lib/modules/*/updates/*
+
+[Action]
+Description = Steamify: NVIDIA modules in the initramfs only if every kernel has them
+When = PostTransaction
+Exec = @SCRIPT@ sync

patches/steamify-nvidia-initramfs.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+# Installed by Steamify CachyOS (SteamOS conversion, NVIDIA fix), run by the
+# pacman hook steamify-nvidia-initramfs.hook before the initramfs is built.
+# The NVIDIA modules go into the initramfs (early KMS) only while EVERY
+# installed kernel has them: mkinitcpio fails on a module a kernel lacks, and
+# limine-mkinitcpio then skips that kernel's boot entry, kernel parameters
+# included. Decided again at every kernel or driver change, so a new kernel
+# without the modules (another kernel, a failed DKMS build, a downgrade) gets
+# a working initramfs and the file comes back once it has them.
+#   steamify-nvidia-initramfs [sync|check]
+#   sync   (default) write or remove the drop-in to match
+#   check  exit status 0 when every installed kernel has nvidia_drm
+conf="${NVIDIA_INITRAMFS_CONF:-/etc/mkinitcpio.conf.d/90-steamify-nvidia.conf}"
+mods="${NVIDIA_MODULES_DIR:-/usr/lib/modules}"
+
+everywhere() {
+    local k found=0
+    for k in "$mods"/*/; do
+        [[ -f "$k/vmlinuz" || -f "$k/pkgbase" ]] || continue   # an installed kernel, not a leftover folder
+        find "$k" -name 'nvidia-drm.ko*' -print -quit 2>/dev/null | grep -q . || return 1
+        found=1
+    done
+    [[ $found == 1 ]]
+}
+
+case "${1:-sync}" in
+    check) everywhere ;;
+    sync)
+        if everywhere; then
+            [[ -f "$conf" ]] || { mkdir -p "$(dirname "$conf")" &&
+                echo 'MODULES+=(nvidia nvidia_modeset nvidia_uvm nvidia_drm)' > "$conf"; }
+        else
+            rm -f "$conf"
+        fi ;;
+    *) echo "usage: $0 [sync|check]" >&2; exit 2 ;;
+esac

tests/nvidia-test.sh

@@ -7,7 +7,7 @@ T="$(mktemp -d)"; trap 'rm -rf "$T"' EXIT
 fail=0
 check() { if eval "$2"; then echo "ok   $1"; else echo "FAIL $1"; fail=1; fi; }
 
-source lib/common.sh; source lib/hdmi-refresh.sh; source lib/nvidia.sh
+SCRIPT_DIR="$PWD"; source lib/common.sh; source lib/hdmi-refresh.sh; source lib/nvidia.sh
 sudo() { "$@"; }
 info() { :; }; ok() { :; }; warn() { :; }; err() { echo "ERR $*"; }
 backup_file() { [[ -f "$1.bak" ]] || cp "$1" "$1.bak"; }
@@ -19,7 +19,8 @@ echo 0x8086 > "$T/drm/card0/device/vendor"; echo 0x030000 > "$T/drm/card0/device
 echo 0x10de > "$T/drm/card1/device/vendor"; echo 0x030000 > "$T/drm/card1/device/class"
 echo 0x2c02 > "$T/drm/card1/device/device"
 NVIDIA_DRM_DIR="$T/drm"
-mkdir -p "$T/mods/6.1.0-cachyos/kernel"; : > "$T/mods/6.1.0-cachyos/kernel/nvidia-drm.ko.zst"
+mkdir -p "$T/mods/6.1.0-cachyos/kernel"; : > "$T/mods/6.1.0-cachyos/kernel/nvidia-drm.ko.zst"; : > "$T/mods/6.1.0-cachyos/pkgbase"
+NVIDIA_SCRIPT="$T/libexec/steamify-nvidia-initramfs"; NVIDIA_HOOK="$T/hooks/85-steamify-nvidia-initramfs.hook"
 NVIDIA_MODULES_DIR="$T/mods"
 NVIDIA_LIMINE_CONF="$T/none/limine.conf"; NVIDIA_SDBOOT_DIR="$T/none/entries"; NVIDIA_GRUB_CFG="$T/none/grub.cfg"   # not generated here: not checked
 
@@ -39,15 +40,27 @@ for loader in limine sdboot grub; do
     nvidia_enable
     check "$loader: parameters added" 'grep -q "nvidia-drm.modeset=1" "$F" && grep -q "nvidia-drm.fbdev=1" "$F"'
     check "$loader: initramfs drop-in written" 'grep -q nvidia_drm "$NVIDIA_INITRAMFS_CONF"'
+    check "$loader: script and pacman hook installed" '[[ -x "$NVIDIA_SCRIPT" ]] && grep -q "^Exec = $NVIDIA_SCRIPT sync" "$NVIDIA_HOOK"'
     check "$loader: backup made, rebuilt once" '[[ -f "$F.bak" && $REBUILDS == 1 ]]'
     SUM="$(cat "$F")"; nvidia_enable
     check "$loader: second run changes nothing" '[[ "$(cat "$F")" == "$SUM" && $REBUILDS == 1 ]]'
     nvidia_disable
-    check "$loader: disable restores the file and drops the drop-in" '[[ "$(cat "$F")" == "$ORIG" && ! -f "$NVIDIA_INITRAMFS_CONF" ]]'
+    check "$loader: disable restores the file, drops the drop-in, script and hook" '[[ "$(cat "$F")" == "$ORIG" && ! -f "$NVIDIA_INITRAMFS_CONF" && ! -e "$NVIDIA_SCRIPT" && ! -e "$NVIDIA_HOOK" ]]'
 done
 
+# The pacman hook's script decides again at every kernel change.
+syncit() { NVIDIA_INITRAMFS_CONF="$NVIDIA_INITRAMFS_CONF" NVIDIA_MODULES_DIR="$NVIDIA_MODULES_DIR" bash "$NVIDIA_SCRIPT" sync; }
+nvidia_hook_install; rm -f "$NVIDIA_INITRAMFS_CONF"
+syncit; check "hook script: every kernel has the modules -> drop-in written" '[[ -f "$NVIDIA_INITRAMFS_CONF" ]]'
+mkdir -p "$T/mods/6.6.0-lts/kernel"; : > "$T/mods/6.6.0-lts/pkgbase"
+syncit; check "hook script: a new kernel without modules -> drop-in removed" '[[ ! -f "$NVIDIA_INITRAMFS_CONF" ]]'
+: > "$T/mods/6.6.0-lts/kernel/nvidia-drm.ko.zst"
+syncit; check "hook script: the modules arrive (DKMS built) -> drop-in back" '[[ -f "$NVIDIA_INITRAMFS_CONF" ]]'
+rm -rf "$T/mods/6.6.0-lts"; mkdir -p "$T/mods/5.0.0-leftover"; : > "$T/mods/5.0.0-leftover/modules.dep"
+syncit; check "hook script: a leftover folder of a removed kernel is ignored" '[[ -f "$NVIDIA_INITRAMFS_CONF" ]]'
+rm -rf "$T/mods/5.0.0-leftover"; nvidia_disable
 # A second kernel without the NVIDIA modules: no early-load drop-in, parameters still set.
-mkdir -p "$T/mods/6.6.0-lts/kernel"; F="$T/sdboot-manage.conf"; printf 'LINUX_OPTIONS="quiet"\n' > "$F"
+mkdir -p "$T/mods/6.6.0-lts/kernel"; : > "$T/mods/6.6.0-lts/pkgbase"; F="$T/sdboot-manage.conf"; printf 'LINUX_OPTIONS="quiet"\n' > "$F"
 nvidia_boot_file() { echo "$F"; }; rm -f "$NVIDIA_INITRAMFS_CONF"; REBUILDS=0; nvidia_enable
 check "kernel without nvidia modules: no drop-in, parameters set" '[[ ! -f "$NVIDIA_INITRAMFS_CONF" ]] && grep -q "nvidia-drm.fbdev=1" "$F"'
 rm -rf "$T/mods/6.6.0-lts"