595.80 (open): kernel oops in nvEvoDisableVblankSemControl on hibernate entry — NULL page fault, GPU left hard-locked (krcWatchdog loop)

Summary

Attempting to hibernate (systemctl hibernate) on a hybrid-graphics laptop with the
dGPU active crashes in the nvidia PM notifier before the hibernation image is written.
systemd-sleep is killed mid-notifier ("exited with irqs disabled"), hibernation is
aborted, and the GPU is left permanently locked (RC watchdog loop) — the Wayland
compositor never recovers and a hard reboot is required. 100% reproducible on first
attempt under the conditions below.

Environment

• Driver: 595.80, open kernel module (Fedora RPMFusion akmod-nvidia-595.80-1.fc44,
  kmod-nvidia-7.0.13-200.fc44.x86_64-595.80-1.fc44)
• Kernel: 7.0.13-200.fc44.x86_64 (Fedora 44), PREEMPT(lazy); Secure Boot disabled,
  lockdown [none]
• GPU: RTX 4060 Laptop (Max-Q), PCI 01:00.0, hybrid with AMD Raphael iGPU (09:00.0)
• Machine: ASUS ROG Strix G713PV (BIOS G713PV.327), Ryzen 9 7845HX, 64 GB RAM
• Session: KDE Plasma 6.7.1, Wayland (kwin_wayland compositing on the dGPU-driven
  external output)
• Displays: internal eDP on the AMD iGPU; external HDMI monitor driven by the NVIDIA
  dGPU, connected and active at the time of the crash
• Module options: NVreg_PreserveVideoMemoryAllocations=1,
  NVreg_TemporaryFilePath=/var/tmp, NVreg_DynamicPowerManagement=0x00
• nvidia-suspend/resume/hibernate systemd services: enabled
• Hibernate target: 59.6 GiB disk swap partition (zram excluded), resume=UUID=... on
  the kernel cmdline, initramfs rebuilt with resume support; ~14 GiB RAM in use
• Platform sleep: firmware exposes s2idle only (no S3)

Steps to reproduce

1. Hybrid mode (dGPU active, external HDMI display connected), Wayland session running.
2. systemctl hibernate

Expected

Hibernation image written; machine powers off; resume restores the session (or, if
something is unsupported, a clean refusal with no side effects).

Actual

Kernel oops inside the nvidia PM notifier during hibernate entry — before the image
write starts:

BUG: unable to handle page fault for address: 00000000000026b0
Oops: Oops: 0000 [#1] SMP NOPTI
CPU: 13 UID: 0 PID: 123864 Comm: systemd-sleep Tainted: G OE 7.0.13-200.fc44.x86_64 #1 PREEMPT(lazy)
Hardware name: ASUSTeK COMPUTER INC. ROG Strix G713PV_G713PV/G713PV, BIOS G713PV.327 08/23/2023
RIP: 0010:nvEvoDisableVblankSemControl+0x34/0xd0 [nvidia_modeset]
Call Trace:
 <TASK>
 nv_set_system_power_state.part.0+0x72/0x1b0 [nvidia]
 nv_pm_notifier+0x35/0xa0 [nvidia]
 notifier_call_chain+0x70/0x110
 blocking_notifier_call_chain_robust+0x73/0xf0
 pm_notifier_call_chain_robust+0x1d/0x40
 hibernate+0x10d/0x190
 state_store+0x45/0x60
 kernfs_fop_write_iter+0x171/0x220
 vfs_write+0x281/0x570
 ksys_write+0x7b/0x110
 do_syscall_64+0x109/0x6e0
note: systemd-sleep[123864] exited with irqs disabled

The fault address 0x26b0 is a small offset from NULL — this looks like a NULL/stale
device pointer dereferenced while disabling vblank semaphore controls in the hibernate
prepare path.

Aftermath: systemd-hibernate.service: Failed with result 'signal'; the kernel then
logs the RC watchdog indefinitely (~every 8 s):

NVRM: krcWatchdog_IMPL: RC watchdog: GPU is probably locked!  Notify Timeout Seconds: 7

The Wayland compositor on the dGPU output never recovers; hard reboot required.

Notes

• Suspend (s2idle) on this configuration has a separate long-standing fragility
  (VRAM corruption / Xid 31 on interrupted suspend), so hibernate was the intended
  supported sleep path here.
• This was the first hibernate attempt on this install; reproduced immediately.
• Happy to provide nvidia-bug-report.log.gz, test patches, or try newer drivers.

Possibly related (different symbol/path, also 595.x mobile PM): #1142 (nv_suspend_devices WARNING on suspend). This report is a hard oops in nvEvoDisableVblankSemControl on the hibernate prepare path specifically.

Start with the reported hibernate call chain: nv_pm_notifier, nv_set_system_power_state, and nvEvoDisableVblankSemControl in the nvidia and nvidia_modeset modules. Reproduce with the stated hybrid-graphics, Wayland, and external-HDMI setup, then trace the failing notifier path and compare it with issue #1142. Done means hibernate either completes or refuses cleanly without an oops, GPU lock, or compositor failure.