RTX 5090 (GB202): GNOME Wayland suspend never reaches S3 — gnome-shell stays runnable in nvidia.ko, nvidia-sleep.sh deadlocks (595.84 open, UseKernelSuspendNotifiers=0)

NVIDIA Open GPU Kernel Modules Version

595.84 (Ubuntu package nvidia-driver-595-open / nvidia-kernel-common-595 595.84-0ubuntu0.24.04.1)

NVRM: loading NVIDIA UNIX Open Kernel Module for x86_64 595.84

Proprietary driver comparison

RTX 5090 / GB202 requires the open kernel modules on this platform. The proprietary-only package is not a viable alternative test. Happy to retest a build NVIDIA specifies.

This hang is inside nvidia.ko (gvaspaceAlloc_IMPL / dmaAllocMapping_GM107 / nv_procfs_write_suspend), not in a userspace compositor bug alone.

Operating System and Version

Ubuntu 24.04.4 LTS (Noble)

Kernel Release

Linux redhead 7.0.0-28-generic #28~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC (Ubuntu HWE, stable, not -rc)

Hardware: GPU

GPU 0: NVIDIA GeForce RTX 5090 (UUID: GPU-d695c932-045f-c37d-e2b6-9bc983de4386) PCI 0000:01:00.0

• Board: MSI MS-7D89, Intel i9-14900K, 188 GiB RAM
• Display: LG 3440x1440 on NVIDIA HDMI (card2-HDMI-A-5)
• Intel UHD 770 iGPU is present but unused for display
• Session: GNOME 46 / Wayland (ubuntu-wayland)

Driver PM config

PreserveVideoMemoryAllocations: 1
UseKernelSuspendNotifiers: 0
EnableS0ixPowerManagement: 0
TemporaryFilePath: /var
mem_sleep: s2idle [deep]
nvidia-drm.modeset=1
nvidia-suspend.service / nvidia-resume.service: enabled

/var has ~748G free (VRAM dump space is not the limiter).

Describe the bug

On GNOME Wayland, systemctl suspend / the GNOME menu does not enter S3. The display goes black, fans and case lighting stay on, keyboard and power button do nothing. Hard power-off is required.

This is the userspace nvidia-sleep.sh path (UseKernelSuspendNotifiers=0), not the kernel PM-notifier / GSP-unload path in #1284.

Failed cycle (2026-08-12 22:53):

1. nvidia-suspend.service runs nvidia-sleep.sh suspend (including chvt 63 + write to /proc/driver/nvidia/suspend). NVRM assertions fire in gmmu_walk.c / mmu_walk_sparse.c during that write.
2. PM: suspend entry (deep)
3. Userspace freeze fails after 20.000s: gnome-shell state:R, stack in NVIDIA (gvaspaceAlloc_IMPL → dmaAllocMapping_GM107 → kbusMapFbAperture_GM107 → memMap_IMPL).
4. Kernel retries s2idle; freeze fails again (2 tasks).
5. Failed to put system to sleep. System resumed again: Device or resource busy
6. nvidia-sleep.sh then blocks >122s in nv_procfs_write_suspend; kernel reports it is blocked on an rw-semaphore owned by gnome-shell.
7. systemd SIGKILLs nvidia-sleep.sh. GPU reload fails:

NVRM: nvAssertOkFailedNoLog: ... NV_ERR_GENERIC ... gpuStateLoad(...) @ gpu_suspend.c:281

8. Display stays off; machine stays fully powered. A later 07:01 retry starts from this already-broken GPU (Flip event timeout on head 0, Failed to allocate NVKMS memory for GEM object) and ends at PM: suspend entry (deep) with no further kernel logs until a hard reset the next morning.

Confirmed workaround (not a driver fix): SIGSTOP gnome-shell before nvidia-suspend.service, SIGCONT after nvidia-resume.service.

With that, the same machine:

• freeze completes in 0.001s
• enters ACPI S3 (Preparing to enter system sleep state S3 / Waking up from system sleep state S3)
• power LED blinks
• power-button press resumes to the lock screen
• GNOME menu path and systemctl suspend both succeed (tested 2026-08-13)

So the GPU can suspend/resume on this board/BIOS/S3 path. The failure is gnome-shell remaining runnable inside NVIDIA ioctl/RM code across the freezer + procfs suspend handshake.

Bypassing systemd (rtcwake -m mem, no nvidia-sleep.sh) fails immediately, as documented:

NVRM: 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 returns -5

To Reproduce

1. Ubuntu 24.04 + nvidia-driver-595-open 595.84 on RTX 5090.
2. Enable NVreg_PreserveVideoMemoryAllocations=1 (required for GNOME Wayland).
3. Leave UseKernelSuspendNotifiers=0 (Ubuntu default; nvidia-suspend.service enabled).
4. Log into GNOME Wayland with the NVIDIA card driving the display.
5. Suspend from the system menu or systemctl suspend.
6. Observe: display off, fans on, no S3, no wake. Journal matches the freeze / nv_procfs_write_suspend deadlock above.

Precondition: do not SIGSTOP gnome-shell first.

Bug Incidence

Always (menu suspend and systemctl suspend) until gnome-shell is paused. After the pause workaround: 2/2 successful S3 + resume cycles, including the GNOME menu + power-button wake.

Related issues

• #472 — PreserveVideoMemoryAllocations without procfs interface (same -5 if systemd units are skipped)
• #1117 — Blackwell + Linux 7.0 s2idle hang (different path; we use deep / S3)
• #1284 — RTX 5090 suspend Xid 120 with UseKernelSuspendNotifiers=1 on 610.x (we are on 595.84 userspace-sleep path)
• Ubuntu already ships SYSTEMD_SLEEP_FREEZE_USER_SESSIONS=false in nvidia-kernel-common to avoid the Xorg/VT-switch freezer deadlock. The kernel freezer still hits this Wayland/gnome-shell case.
• GNOME: VT switch in nvidia-sleep.sh is Xorg-oriented (gnome-settings-daemon#903)

What a real driver fix would look like

gnome-shell (and any Wayland compositor) issuing NVIDIA ioctls during nvidia-sleep.sh / freezer should be interruptible / freezer-aware. Today those threads stay R in RM (gvaspaceAlloc / BAR1 map) and then deadlock with nv_procfs_write_suspend. SIGSTOP is only a userspace bandage.

Attachments

See gist (journals, systemd workaround units, nvidia-bug-report.log):

https://gist.github.com/robertvandervoort/5488a63af729e7a1f474d61c28f49653

Start with the nvidia-sleep.sh suspend path and the nv_procfs_write_suspend call, then compare the reported gnome-shell stack with the related suspend behavior in #1284. Trace how the freezer interacts with the NVIDIA RM mappings and procfs suspend handshake. Done means suspend reaches S3 and resumes reliably without SIGSTOPing gnome-shell or leaving the GPU in a broken state.