libnvidia-hide

Prevent Electron apps from waking NVIDIA dGPU on hybrid laptops (LD_PRELOAD)

---

What this is

libnvidia-hide is a userspace NVIDIA-hiding solution for hybrid GPU laptops.

It consists of:

• a shared library (libnvidia-hide.so) that hides NVIDIA devices from a process using LD_PRELOAD
• a launcher binary (nvidia-hide) that applies the library per application, without requiring a global preload

The goal is to prevent Electron / Chromium-based applications (VS Code, Discord, Slack, etc.) from waking the NVIDIA dGPU unnecessarily, while keeping:

• full filesystem access
• portals, camera, microphone, clipboard
• normal desktop integration

No sandboxing, cgroups, udev rules, kernel patches, or driver removal are involved.

---

The problem

On modern hybrid GPU laptops, Electron applications frequently:

• enumerate all DRM devices (/dev/dri/card*, renderD*)
• load NVIDIA GBM / GLX / Vulkan libraries
• parse NVIDIA Vulkan ICDs and implicit layers
• probe PCI configuration space for GPUs
• open /dev/nvidia* character devices

Any of the above is enough to runtime-resume the NVIDIA GPU, causing:

• unnecessary power usage
• slow application startup
• possibly broken power management

Environment variables such as:

• DRI_PRIME=0
• __GLX_VENDOR_LIBRARY_NAME=mesa
• Vulkan loader overrides

are not sufficient in many real-world cases.

---

What this project does

libnvidia-hide prevents NVIDIA from being considered at discovery time, not just at usage time.

1. Dynamic NVIDIA detection

• Scans /sys/class/drm/*/device/vendor
• Identifies NVIDIA DRM nodes by vendor ID (0x10de)
• Resolves corresponding PCI BDFs dynamically
• No hardcoded card numbers or assumptions

2. Hides NVIDIA from filesystem enumeration

Filters NVIDIA-related entries from:

• /dev
• /dev/dri
• /dev/dri/by-path

As a result, Electron never “sees” NVIDIA devices during probing.

3. Blocks NVIDIA device access

Prevents access to:

• NVIDIA renderD* nodes
• /dev/nvidia* character devices

4. Blocks NVIDIA userspace stacks

Prevents loading of:

• libGLX_nvidia.so
• nvidia-drm_gbm.so
• libnvidia-*

This stops Chromium / Electron from selecting NVIDIA paths early.

5. Prevents PCI-level probing

Blocks reads of:

• /sys/.../<NVIDIA_BDF>/config

This avoids runtime PM wakeups even when character devices are blocked.

6. Works with Electron’s multi-process model

• Every Electron subprocess loads the library
• Initialization happens once per process
• Repeated debug output is expected and correct

---

What this project deliberately does not do

• No sandboxing
• No cgroup device filtering
• No permission or group changes
• No driver blacklisting
• No kernel patches
• No system-wide preload unless you explicitly choose to do so

---

Limitations

• Does not work with Flatpak / Snap applications
  • LD_PRELOAD is blocked by design in sandboxed environments

---

Building

This repository from now on ships source only. You are expected to build locally.

Requirements

• gcc
• glibc (for libdl)

Build commands

gcc -O2 -fPIC -Wall -Wextra -std=c11 \
  -shared -ldl \
  -o libnvidia-hide.so libnvidia-hide.c


gcc -O2 -Wall -Wextra -std=c11 \
  -o nvidia-hide nvidia-hide.c

(Optional) install locations:

install -Dm755 libnvidia-hide.so /usr/local/lib/libnvidia-hide.so
install -Dm755 nvidia-hide /usr/local/bin/nvidia-hide

Alternatively, you can use the included makefile, and run:

make
sudo make install

---

How to use

Recommended: launcher-based usage (no global preload)

nvidia-hide run -- code

This:

• sets LD_PRELOAD only for that process tree
• automatically applies policy (allowlist / denylist)
• avoids polluting your entire desktop session

---

Optional: manual LD_PRELOAD usage

If you want to preload manually:

LD_PRELOAD=/path/to/libnvidia-hide.so code

---

Configuration: allowlist / denylist

The library decides whether it should be active per process by inspecting /proc/self/exe.

Config files (recommended)

Location:

~/.config/nvidia-hide/allowlist
~/.config/nvidia-hide/denylist

Format:

• one glob pattern per line
• # comments supported
• empty lines ignored

Matching rules:

• patterns without / match the executable basename
• patterns with / match the full executable path

Examples:

# allow only VS Code
code
*/visual-studio-code/code

# never hide NVIDIA from these
grep
bash

# allow discord
echo "discord" > ~/.config/nvidia-hide/allowlist

# deny discord
echo "discord" > ~/.config/nvidia-hide/denylist 

Precedence rules

1. If an allowlist exists, the library is inactive unless matched
2. Denylist always wins

If policy results in active=0, the library becomes a true no-op:

• no DRM probing
• no sysfs scanning
• no side effects

---

Environment-based configuration (optional)

Instead of files, you may use env vars (colon-separated globs):

LIBNVIDIAHIDE_ALLOWLIST="code:electron*"
LIBNVIDIAHIDE_DENYLIST="bash:grep"

---

Debugging

Enable verbose logging:

LIBNVIDIAHIDE_DEBUG=1 nvidia-hide run -- code

Example output:

[libnvidia-hide] policy: exe=/opt/visual-studio-code/code
[libnvidia-hide] policy: active=1 (has_allow=1 allow_match=1 deny_match=0)
[libnvidia-hide] init: nvidia_nodes=2 nvidia_bdfs=1
[libnvidia-hide]   node: card1
[libnvidia-hide]   node: renderD129
[libnvidia-hide]   bdf:  0000:01:00.0

---

Verifying that the dGPU stays asleep

Runtime PM status

watch -n0.5 cat /sys/bus/pci/devices/0000:01:00.0/power/runtime_status

Expected:

suspended

Tracing NVIDIA opens

sudo bpftrace -e '
tracepoint:syscalls:sys_enter_openat
/str(args->filename) ~ "nvidia|renderD"/
{ printf("%s %d %s\n", comm, pid, str(args->filename)); }'

With libnvidia-hide active, no NVIDIA paths should appear.

---

Final notes

This is a userspace-only workaround:

• no kernel patches
• no driver removal
• no power-management hacks

You essentially force electron to behave like a good citizen.