RTX 5090/GB202: deep S3 resume wedges GSP/NVKMS and removes GPU from PCIe until AC power loss

Graphics / Linux Linux
1

RTX 5090/GB202: deep S3 resume wedges GSP/NVKMS, blanks displays and can remove GPU from PCIe until AC power is disconnected

System

  • GPU: Gigabyte GeForce RTX 5090, GB202, PCI ID 10de:2b85, normally at 0000:01:00.0
  • Motherboard: ASUS ROG STRIX X670E-E GAMING WIFI
  • BIOS: 3902
  • CPU/platform: AMD AM5/Raphael with integrated AMD graphics
  • OS: Ubuntu 26.04.1 LTS
  • Kernel: 7.0.0-30-generic
  • Desktop: GNOME/Wayland
  • Displays: two Samsung Odyssey monitors connected directly to the RTX 5090 over DisplayPort; a small HDMI display is connected to the AMD iGPU
  • NVIDIA package: Ubuntu nvidia-driver-610-open 610.43.02-0ubuntu0.26.04.1
  • NVIDIA kernel/GSP/userspace versions: aligned at 610.43.02
  • Sleep modes: s2idle [deep]; deep is selected
  • Secure Boot was enabled with the Ubuntu MOK-enrolled module during the first failures. The failure also reproduced after Secure Boot was disabled.
  • Resizable BAR: enabled, BAR1 assigned the full 32 GiB aperture during healthy boots
  • NVreg_PreserveVideoMemoryAllocations=1
  • NVreg_UseKernelSuspendNotifiers=0
  • NVreg_EnableS0ixPowerManagement=0
  • Platform reports NVIDIA S0ix support as not supported
  • nvidia-suspend.service, nvidia-resume.service, and nvidia-hibernate.service enabled

Reproduction

  1. Cold boot; verify RTX 5090 at 01:00.0, both DP displays active, and nvidia-smi healthy.
  2. Request deep S3 through systemd.
  3. Wake using a Wake-on-LAN magic packet from another Linux host.
  4. The machine powers/resumes, but the two RTX 5090 displays remain black and the graphical desktop does not recover.
  5. A warm reboot may leave the RTX 5090 entirely absent from lspci; the NVIDIA module repeatedly logs NVRM: No NVIDIA GPU found.
  6. Shut down, switch the PSU off, disconnect mains, discharge the machine, and cold boot. The RTX 5090 enumerates again at 01:00.0 and operates normally.

Observed failures

The first visible failure repeatedly printed:

nvidia-modeset: ERROR: GPU:0: Error while waiting for GPU progress
INFO: task plymouthd blocked for more than 122 seconds
INFO: task plymouthd is blocked on a mutex likely owned by task gnome-shell

reboot, restart, and exit typed at that console could not recover the wedged system. A hard reset was required.

In later controlled cycles, the persistent journal records PM: suspend entry (deep) but no PM: suspend exit before the forced reset. Following a warm boot in the latched state, the RTX 5090 does not appear in lspci, no NVIDIA DRM card/connectors exist, nvidia-smi returns code 9, and the kernel repeatedly reports:

NVRM: No NVIDIA GPU found.

The AMD iGPU remains operational, allowing diagnosis from its HDMI display.

Isolation already performed

  • Updated motherboard BIOS from 3205 to 3902.
  • Updated from NVIDIA 595 open to Ubuntu NVIDIA 610 open; MOK enrollment completed and all module/userspace/GSP versions verified aligned.
  • Verified Wake-on-LAN mode magic, PCI wake enabled, and successful magic-packet delivery.
  • Closed CLU VIEW and confirmed its lhm_reader process released every /dev/nvidia* descriptor before suspend. Failure remained.
  • Paused the sole gnome-shell process with SIGSTOP before the NVIDIA suspend service path, verified it entered stopped state, and attempted deep S3 under a fail-closed transaction supervised by a second Linux host. The displays still did not recover.
  • Tested with Secure Boot enabled and later disabled. Failure remained.
  • Confirmed the GPU, cabling and both DP outputs work normally after a true AC power removal.
  • The system now masks suspend/hibernate targets to prevent further accidental failures.

Expected result

Deep S3 should resume the RTX 5090 and both attached displays, or the driver should fail the suspend transaction cleanly without wedging the GPU/GSP or leaving the PCIe function absent across a warm reboot.

Related reports

  • NVIDIA/open-gpu-kernel-modules #1284: RTX 5090 suspend failure on 610.57.04, deep and s2idle
  • NVIDIA/open-gpu-kernel-modules #1291: RTX 5090 GNOME/Wayland suspend deadlock
  • NVIDIA/open-gpu-kernel-modules #1281: Blackwell nvidia_modeset PM lock deadlock after S3

The attached NVIDIA bug report was captured after a full AC power cycle restored the GPU. Filtered persistent-kernel logs include the failed deep-suspend boots and the subsequent boot where the GPU was absent.

kernel-boot1-filtered-sanitized.log (19.4 KB)

kernel-boot4-filtered-sanitized.log (16.9 KB)

kernel-boot5-filtered-sanitized.log (16.9 KB)

nvidia-bug-report-sanitized.log.gz (278.9 KB)

2

Average Joe user here, so apologize if these are naive questions. But I’ve been struggling with suspend/resume issues on much older hardware+driver configuration lately (GTX 750+580 proprietary driver) and I’ve learned a couple things.

  • Is it only wake-on-LAN that fails to resume properly? Can you resume from the console of the suspended machine?
  • Can you resume the displays and desktop GUI by performing a manual VT switch from the console of the suspended machine? My KDE desktop seems to run on TTY2 and under some circumstances Ctrl+Alt+F3 followed by Ctrl+Alt+F2 brings the desktop back.