Where to report nvidia-utils elogind suspend bug? Complete fix included

[bobemoe]

I found and fixed a critical suspend bug in nvidia-utils but struggling to find where to report it properly.

Reporting Issues I Encountered:

• NVIDIA Developer Forums: Requires business name/industry signup (I'm just a personal user)
• NVIDIA Support pages: Also require business names and extensive signup info for personal users
• This repo's Issues: Says "only for bugs specific to the open kernel driver" but I'm using proprietary although this probably affects both

The Bug:

The /usr/lib/elogind/system-sleep/nvidia script (from nvidia-utils 575.57.08) is missing the "pre" case for regular suspend operations, causing suspend to fail on non-systemd systems using elogind.

Error symptoms:

NVRM: GPU 0000:01:00.0: PreserveVideoMemoryAllocations module parameter is set. 
System Power Management attempted without driver procfs suspend interface.
nvidia 0000:01:00.0: PM: pci_pm_suspend(): nv_pmops_suspend [nvidia] returns -5
PM: Some devices failed to suspend, or early wake event detected

Root Cause:

The current script only has a "post" case:

case "$1" in
    post)
        /usr/bin/nvidia-sleep.sh "resume"
        ;;
esac

But it's missing the "pre" case that should suspend the GPU before system suspend. Interestingly, the script handles suspend-then-hibernate correctly with both pre/post cases, but not regular suspend.

The Fix:

Add the missing "pre" case to /usr/lib/elogind/system-sleep/nvidia. The current script is:

case "$1" in
    post)
        /usr/bin/nvidia-sleep.sh "resume"
        ;;
esac

It should be:

case "$1" in
    pre)
        case "$2" in
            suspend)
                echo suspend > /proc/driver/nvidia/suspend
                ;;
            hibernate)
                echo hibernate > /proc/driver/nvidia/suspend
                ;;
        esac
        ;;
    post)
        /usr/bin/nvidia-sleep.sh "resume"
        ;;
esac

Why This Affects Both Driver Types:

Both nvidia (proprietary) and nvidia-open packages depend on the same nvidia-utils package, which provides this broken script. So this bug likely affects all NVIDIA Linux users on non-systemd systems.

Verification:

• Manual GPU suspend/resume works: echo suspend > /proc/driver/nvidia/suspend
• The fix works: suspend/resume now works perfectly
• Using: Artix Linux (OpenRC/elogind), proprietary driver 575.57.08

Questions:

1. Where should I officially report this?
2. Can anyone with open drivers confirm they have the same issue?
3. Can an NVIDIA dev just fix this in nvidia-utils?

I have a working pacman hook to preserve the fix across updates, but this really should be fixed upstream in the nvidia-utils package.

System details:

• nvidia 575.57.08-5 (proprietary)
• nvidia-utils 575.57.08-3
• Artix Linux with OpenRC/elogind
• GPU: GeForce RTX 3090

[aritger]

Thanks for your report and the proposed fix. I've filed NVIDIA internal bug 5382200 for this.

The /usr/lib/elogind/system-sleep/nvidia script I assume comes from your distribution, but I think it is the same and/or is adapted from systemd/system-sleep/nvidia packaged in the .run file packaging of the NVIDIA GPU driver. I don't know yet whether the missing "pre" is a problem in systemd/system-sleep/nvidia for systemd usage, too, or if this should simply be considered a problem in the Artix Linux adaptation of the script for use with elogind. But, we'll try to investigate a bit deeper.

[aaronp24]

As Andy noted, the systemd/system-sleep/nvidia script is only intended to be used with systemd and isn't designed to work with other init systems. For systemd, the pre case is handled by the nvidia-suspend.service and nvidia-hibernate.service unit files. The suspend-then-hibernate case is special because it cannot be handled correctly by other systemd service files.

As described in the README, it's up to the distribution and its init system to properly bracket suspend & resume, which in this case I think would mean providing a vendor-specific elogind/system-sleep script that handles the pre case rather than just copying the systemd one.

I realize that this stuff gets complicated. I'll see what I can do to simplify things on the driver side if possible.