OPINIONATED OMARCHY
TROUBLESHOOTING CORPUS
790
RECORDS
790
AUDITED
0
UNCHECKED
2388
SOURCES
LIVE
gpu-drivers high very-common DISPUTE

Stop NVIDIA black screens and corruption after suspend/resume

omarchyarchendeavouroscachyosmanjaronvidiawaylandhyprlandlaptopdesktop
CORRECTED An auditor found the fix wrong or incomplete and rewrote it.
AUDIT NOTE

Checked on this workstation, which is Omarchy 4.0.2-1, kernel 7.1.9, nvidia-open-dkms and nvidia-utils 610.57.04-1 on an RTX 3090, and against the cited pages plus NVIDIA's own 610.57.04 power management README. The symptom and the cause are real. Confirmed from the NVIDIA README that video memory preservation "is handled automatically" on the open modules when NVreg_UseKernelSuspendNotifiers=1, and that /proc/driver/nvidia/suspend "is required when using" PreserveVideoMemoryAllocations=1. Confirmed from the Arch wiki that Arch sets these for supported drivers so preserve works out of the box, which makes the record's hand-written /etc/modprobe.d/nvidia-power.conf redundant on Arch and Omarchy alike. Four things are wrong on Omarchy 4. First, sudo mkinitcpio -P does nothing here: confirmed on this machine that /etc/mkinitcpio.d/ holds zero files and that neither linux nor mkinitcpio installs a preset into it, so the command processes no presets and never rebuilds the UKI that Limine boots. The Omarchy 4 command is sudo limine-mkinitcpio, which is what every Omarchy migration under /usr/share/omarchy/migrations/ calls. Second, the advice to leave the sleep services off on 595 and newer is wrong on Omarchy specifically: confirmed on this machine that gpu-screen-recorder 6.0.1-1 owns /usr/lib/modprobe.d/gsr-nvidia.conf setting NVreg_PreserveVideoMemoryAllocations=1, that the package is line 47 of /usr/share/omarchy/install/omarchy-base.packages, that /proc/driver/nvidia/params therefore reports PreserveVideoMemoryAllocations: 1 on this box, and that systemctl is-enabled reports all three sleep units disabled with no nvidia-suspend string anywhere under /usr/share/omarchy. Upstream quattro's install/hardware/nvidia.sh, retrieved today, still does not enable them, so omacom/omarchy PR 8127 is unmerged and the gap is live on 4.0.3. Reading /usr/lib/systemd/system-sleep/nvidia on this machine confirms it only handles the post resume case for a plain suspend, so the save never runs while the restore is attempted. Third, the danger field missed the sharpest consequence: with Preserve=1 on Omarchy's early-KMS layout, hibernate resume fails and the image is discarded, which omacom/omarchy issue 8126 documents with a kernel trace and the Hyprland wiki warns about in general form. Omarchy ships HOOKS+=(resume), confirmed in /etc/mkinitcpio.conf.d/omarchy_resume.conf, so hibernation is configured by default and this is a real data-loss path. Fourth, "/var/tmp must be on a real filesystem (ext4/XFS, not tmpfs)" is too narrow: Omarchy's default root is btrfs and /var/tmp lives on it here, and the real requirement per NVIDIA is unnamed temporary file support plus capacity. Two smaller points: sort /proc/driver/nvidia/params needs no sudo, confirmed by reading it as an unprivileged user, and the record names three sleep units where four exist. On sources: both cited GitHub issues fail to support the record. Read in full, issue 2635 is an Omarchy 3.0 to 3.8 display-wake bug reported on AMD, Intel Iris and NVIDIA alike, whose stated workaround is systemctl restart sddm, and it never mentions video memory preservation or Xid 13. Issue 2112 is the same family on Omarchy 3, with hypridle.conf as the workaround. Neither names a parameter this record sets, so both go in sources_remove and are replaced with the NVIDIA README, omacom/omarchy issue 8126 and PR 8127, and the upstream driver install script. Severity high and frequency very-common both stand and are left alone: the misconfiguration ships by default to every Omarchy NVIDIA machine. NOT exercised: a suspend cycle. This workstation cannot be suspended for this audit, so option A and option B are argued from the driver README, the shipped unit files and the tracker report, not from a resume I watched. The claim that option B is also needed on some open-driver machines comes from one user's A/B in issue 8126 and is reported, not confirmed here.

Merged on 2026-09-13: nvidia-black-screen-external-after-suspend was retired into this record and its page now 404s. Its 2026-09-13 re-audit found it the same problem as this one and this one strictly better, knowing the gpu-screen-recorder trap, splitting the open and 580xx branches and carrying the hibernate data-loss danger, while that record's own fix was a regression on 595 and newer because it told the reader to enable the three nvidia sleep services that upstream disables. The auditor rewrote it to display-side triage that deferred here for the memory decision, which left it no longer a duplicate, so this merge consolidates rather than removing redundancy: its triage steps are now the first section of the fix above, ahead of the driver checks, because they are cheaper and they settle the commonest case. It also carried systemd-boot and grub in applies_to, wrong for the Omarchy branch, and that defect goes with it. Its full text is in the git history and in raw/o3-core-display-audit.json.

The cause above was rewritten on 2026-09-13 to match this note; the fix was corrected by the audit itself.
RISK

Under PreserveVideoMemoryAllocations=1 the driver demands the /proc/driver/nvidia/suspend handshake at every power transition. Omarchy early-loads the NVIDIA modules, so on a hibernate resume the initramfs kernel already has nvidia bound and there is no systemd to drive that handshake. The image loads and is then discarded, with PM: hibernation: Failed to load image, recovering in the log, followed by a cold boot, and every unsaved thing that was open is gone. Omarchy also ships HOOKS+=(resume) in /etc/mkinitcpio.conf.d/omarchy_resume.conf, so hibernation is configured out of the box. Suspend and hibernate are mutually exclusive under this configuration, so decide which one you want before changing anything. The Hyprland wiki carries the same warning in general form: loading the NVIDIA modules early may stop resume from hibernation working, and the machine just boots instead. Separately, /var/tmp has to hold a full copy of video memory plus about 5 percent, so a short root filesystem makes suspend fail or hang. Any change under /etc/modprobe.d needs the initramfs rebuilt because of early KMS, and an interrupted limine-mkinitcpio leaves an unbootable UKI, so do not power-cycle while it runs.

Symptom

The laptop or desktop suspends fine, but on wake the monitors stay black or show 'No Signal', with the machine still running (you can SSH in, or Ctrl+Alt+F2 to a TTY and back sometimes helps), or the desktop comes back visibly corrupted. journalctl -b -1 contains lines like:


NVRM: Xid (PCI:0000:08:00): 13, pid='<unknown>', name=<unknown>, Graphi>
nvidia-modeset: ERROR: GPU:0: Failed detecting connected display devices
nvidia-modeset: WARNING: GPU:0: Unable to read EDID for display device ... (DP-0)

Cause

The NVIDIA driver saves and restores only essential video memory across a suspend cycle by default. The rest is lost, the userspace driver cannot always reconstruct it, and the result is rendering corruption, dead outputs and Xid 13 errors from whatever was drawing. Full preservation is reached two different ways depending on the driver branch. The open kernel modules from 595 onward do it themselves when NVreg_UseKernelSuspendNotifiers=1, which is what Arch ships in /usr/lib/modprobe.d/nvidia-sleep.conf. The 430 to 590 branch, which includes the nvidia-580xx driver Omarchy installs on Maxwell, Pascal and Volta, needs NVreg_PreserveVideoMemoryAllocations=1 plus the nvidia-suspend, nvidia-hibernate and nvidia-resume services, because under that parameter NVIDIA requires the save and restore to be driven through /proc/driver/nvidia/suspend. On Omarchy 4 the two paths get crossed: gpu-screen-recorder is in the base package set and its /usr/lib/modprobe.d/gsr-nvidia.conf turns PreserveVideoMemoryAllocations on for every NVIDIA machine, while the installer never enables the services that parameter then requires.

Fix

**Rule out the display side first.** A dark output after a wake is often not a memory problem, and these checks are free. From another machine over ssh, or from a TTY:


hyprctl -j monitors | jq -r '.[] | "\(.name) \(.width)x\(.height)@\(.refreshRate) dpms=\(.dpmsStatus) disabled=\(.disabled)"'
  • Nothing answers and the machine is unreachable: the driver or the kernel is gone. Carry on below.
  • A monitor is listed with dpms=false: the panel is asleep, not broken. Wake it. On Hyprland 0.56 hyprctl dispatch evaluates its argument as Lua, so a bare dispatcher name fails with a parse error:

  hyprctl dispatch 'hl.dsp.dpms({ action = "on" })'
  • A monitor is listed at 0x0: it came back with an EDID carrying no modes. Omarchy ships a check for exactly that state:

  omarchy-hyprland-monitor-modeless; echo $?   # 0 means an enabled monitor has no modes

Power cycle the monitor or replug the cable, because Hyprland re-reads the EDID on hotplug. If an external head never reappears on a hybrid machine it is wired to a GPU the compositor did not pick, which is an AQ_DRM_DEVICES problem rather than a suspend one.

  • Everything reads correctly and the screen is still black or garbled: it is the driver side. Carry on below.

Then read what the driver is actually doing. This needs no root:


sort /proc/driver/nvidia/params | grep -E 'PreserveVideoMemoryAllocations|UseKernelSuspendNotifiers|TemporaryFilePath'
systemctl is-enabled nvidia-suspend.service nvidia-hibernate.service nvidia-resume.service
pacman -Qs 'nvidia-open-dkms|nvidia-580xx-dkms'

Which branch you are on decides everything below. Omarchy 4 picks the driver for you in install/hardware/nvidia.sh: nvidia-open-dkms on Turing and newer, nvidia-580xx-dkms on Maxwell, Pascal and Volta.

  • **nvidia-open-dkms (595 and newer, 610.57.04 at the time of writing).** The open kernel modules preserve video memory themselves when UseKernelSuspendNotifiers: 1. nvidia-utils already ships that, together with TemporaryFilePath: "/var/tmp", in /usr/lib/modprobe.d/nvidia-sleep.conf. There is nothing to hand-write, and the three sleep services stay off.
  • **nvidia-580xx-dkms (the 430 to 590 branch).** No suspend notifier support. This branch needs PreserveVideoMemoryAllocations=1 plus nvidia-suspend.service, nvidia-hibernate.service and nvidia-resume.service, because NVIDIA's power management README says the /proc/driver/nvidia/suspend mechanism "is required when using this parameter".

**The Omarchy 4 trap.** gpu-screen-recorder is in the base package set (/usr/share/omarchy/install/omarchy-base.packages), so every Omarchy machine has it, and its packaging ships:


/usr/lib/modprobe.d/gsr-nvidia.conf
options nvidia NVreg_PreserveVideoMemoryAllocations=1

That turns the parameter on for every Omarchy NVIDIA machine, including the open-driver ones that did not ask for it. Omarchy's installer never enables the services that parameter requires (grep -r nvidia-suspend /usr/share/omarchy returns nothing, and upstream quattro still does not do it). /usr/lib/systemd/system-sleep/nvidia only covers the post case for a plain suspend, so the restore is attempted on the way up while the save on the way down never ran. That is the asymmetry that produces Xid 13 across the compositor, the shell and the browser on resume.

Pick one of the two consistent configurations and do not leave the machine between them.

**Option A, let the notifier path own it.** Open driver only. Keeps hibernation working. A file in /etc/modprobe.d with the same basename completely replaces the one in /usr/lib/modprobe.d, so shadow the gpu-screen-recorder file:


sudo tee /etc/modprobe.d/gsr-nvidia.conf >/dev/null <<'EOF'
# Shadows /usr/lib/modprobe.d/gsr-nvidia.conf from gpu-screen-recorder.
# The open kernel modules preserve video memory through
# NVreg_UseKernelSuspendNotifiers=1, which nvidia-utils already sets.
options nvidia NVreg_PreserveVideoMemoryAllocations=0
EOF
sudo limine-mkinitcpio

**Option B, honour PreserveVideoMemoryAllocations=1.** Required on nvidia-580xx-dkms, and reported on the tracker to be needed on some open-driver machines too. Enable the units NVIDIA's README asks for:


sudo systemctl enable nvidia-suspend.service nvidia-hibernate.service nvidia-resume.service

Add nvidia-suspend-then-hibernate.service if you use suspend-then-hibernate. Read the danger field before choosing this one: on Omarchy's default early-KMS layout it costs you hibernation.

Then check the backing store. It has to hold a copy of all video memory plus about 5 percent, and it has to support unnamed temporary files. Omarchy's default root is btrfs, which does. /tmp is tmpfs on Omarchy and is the wrong place, which is why Arch points the parameter at /var/tmp:


nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits
df -hT /var/tmp

Reboot and test one suspend cycle.

**Regenerating the initramfs on Omarchy 4.** Omarchy early-loads the NVIDIA modules through /etc/mkinitcpio.conf.d/nvidia.conf, and the modconf hook copies both /etc/modprobe.d and /usr/lib/modprobe.d into the image, so any change to those files needs a rebuild. mkinitcpio -P does nothing on Omarchy 4: /etc/mkinitcpio.d/ is empty and the linux package ships no preset. Omarchy boots a UKI through Limine and the command that rebuilds it is:


sudo limine-mkinitcpio

On plain Arch, where /etc/mkinitcpio.d/linux.preset exists, sudo mkinitcpio -P is still the right command.

**Reading the journal from the right boot.** Use journalctl -b -k while the session is still up, and journalctl -b -1 -k only when the machine had to be power cycled:


journalctl -b -k | grep -iE 'Xid \(PCI|nvidia-modeset: ERROR|Failed detecting connected display'

**On Omarchy 4 there is no GRUB and no systemd-boot.** It boots a unified kernel image through Limine. Kernel command line changes go in a drop-in under /etc/limine-entry-tool.d/, /etc/default/grub does not exist, and sudo mkinitcpio -P aborts because /etc/mkinitcpio.d/ holds no presets. The rebuild command is sudo limine-mkinitcpio. On plain Arch with systemd-boot or GRUB the older instructions apply and sudo mkinitcpio -P is correct there.

Verify

Read the parameters back, no root needed, and confirm they agree with the option you chose:


sort /proc/driver/nvidia/params | grep -E 'Preserve|UseKernelSuspend|TemporaryFilePath'
systemctl is-enabled nvidia-suspend.service nvidia-resume.service

The three sleep units must be enabled if and only if PreserveVideoMemoryAllocations: 1. Then suspend once, wake, and read the boot you suspended in:


journalctl -b -k | grep -iE 'Xid \(PCI|nvidia-modeset: ERROR'

That should be empty and the desktop should come back with no corruption. Use journalctl -b -1 -k instead only when the machine had to be power-cycled. If you chose option B, confirm the units actually fired:


journalctl -b -u nvidia-suspend.service -u nvidia-resume.service

Sources

nvidia-suspend-resume-black-screen-vram