Fix suspend failing outright on an NVIDIA machine
The symptom, the NVRM message and the general cause are real, and NVreg_UseKernelSuspendNotifiers genuinely exists (nv-reg.h: "If enabled, this option prompts the NVIDIA kernel module to register a notifier that saves and restores all video memory allocations across system power management cycles if PreserveVideoMemoryAllocations is enabled. 0: Suspend notifiers are not used (default), 1: Suspend notifiers are used when available"). But the '430-590 vs 595+' version boundary is fabricated precision and the instruction to DISABLE the three services on '595+' is actively harmful. The current driver README (610.57.04, matching Arch's nvidia-utils 610.57.04-1) says the notifier path applies "When the open kernel modules are in use" and still documents nvidia-suspend/hibernate/resume as installed and enabled by default. The discriminator is open vs proprietary modules, not a version cut-off. Blindly masking the services on a proprietary-driver box will break suspend rather than fix it.
If you use early KMS (the nvidia module loaded from the initramfs) the driver has no access to NVreg_TemporaryFilePath, so hibernation cannot preserve VRAM. Do not use early KMS if you need hibernation. Changing modprobe options requires regenerating the initramfs or the setting silently does not apply.
Symptom
Suspend aborts and the machine stays awake (or wakes instantly). journalctl -k shows:
NVRM: GPU 0000:01:00.0: PreserveVideoMemoryAllocations module parameter is set. System Power Management attempted without driver procfs suspend interface. Please refer to the 'Configuring Power Management Support' section in the driver README.
PM: pci_pm_suspend(): nv_pmops_suspend+0x0/0x20 [nvidia] returns -5
nvidia 0000:01:00.0: PM: failed to suspend async: error -5
Cause
The NVIDIA driver is set to preserve all video memory across suspend, but the mechanism that actually saves it is not active. There are two such mechanisms and the discriminator is **which kernel modules are in use, not a driver version cut-off**: with the open kernel modules the driver registers a suspend notifier, enabled by NVreg_UseKernelSuspendNotifiers=1. Otherwise it is the nvidia-suspend / nvidia-hibernate / nvidia-resume systemd services, which the current driver README still documents as installed and enabled by default. Disabling those services on a proprietary-module machine breaks suspend rather than fixing it.
Fix
Find out which module flavour you are on. That, not the version number, decides the mechanism:
nvidia-smi --query-gpu=driver_version --format=csv,noheader
modinfo -F license nvidia # "Dual MIT/GPL" = open modules; "NVIDIA" = proprietary
pacman -Qs 'nvidia-open|^nvidia '
sudo sort /proc/driver/nvidia/params | grep -E 'UseKernelSuspendNotifiers|PreserveVideoMemoryAllocations|TemporaryFilePath'
systemctl is-enabled nvidia-suspend.service nvidia-hibernate.service nvidia-resume.service
**Proprietary modules (nvidia / nvidia-dkms): the default, and the case for most users.** The three services ARE the mechanism. They drive /proc/driver/nvidia/suspend. They ship enabled. If you disabled them, put them back:
sudo systemctl enable nvidia-suspend.service nvidia-hibernate.service nvidia-resume.service
Do NOT disable them and do NOT set NVreg_UseKernelSuspendNotifiers here.
**Open kernel modules (nvidia-open / nvidia-open-dkms).** The kernel suspend notifier can handle it instead:
# /etc/modprobe.d/nvidia-power.conf
options nvidia NVreg_PreserveVideoMemoryAllocations=1 NVreg_UseKernelSuspendNotifiers=1 NVreg_TemporaryFilePath=/var/tmp
Only then, and only after confirming suspend works, is it safe to drop the services. Change one thing at a time and test a suspend cycle between each.
Either way, rebuild the initramfs if the nvidia modules are in it:
sudo mkinitcpio -P # or sudo limine-mkinitcpio on Omarchy
The VRAM dump target must support unnamed temporary files (O_TMPFILE, which ext4, XFS and Btrfs all do) and must not be a tmpfs, which is why /var/tmp is used instead of the default /tmp. It needs room for the total VRAM in use:
nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits
df -h /var/tmp
Verify
sudo sort /proc/driver/nvidia/params shows UseKernelSuspendNotifiers: 1 (or PreserveVideoMemoryAllocations: 1 on older drivers) and TemporaryFilePath: "/var/tmp". systemctl suspend now completes without the error -5 lines.