Bug Report: Regression: Multi-GPU KWin Wayland Hard Freeze / Coredump on Cursor Border Crossing (NVIDIA 580.x Drivers / Kernel 7.0 / Quadro)

A catastrophic user-space freeze / KWin coredump occurs on an asymmetric multi-GPU laptop/mobile workstation configuration the exact millisecond the hardware mouse cursor approaches or crosses the screen boundary onto an external monitor.

First note that on Kubuntu Impish Indie with 470 drivers it worked perfectly, 3 external monitors connected via different ways (HDMI / DisplayPort on dock or USB-C to HDMI or all on dock, all configurations successful) with no cursor freeze, everything worked correctly.

It is the upgrade to 580 drivers on (K)Ubuntu 26.04 LTS that caused the problem.

Hardware Profile:

  • Device: HP ZBook Studio x360 G5 Mobile Workstation with Xeon E-2186M
  • Integrated GPU: Intel UHD Graphics P630 (Drives internal 4K Panel, 200% Scale, 10-bit color profile). Scaling and color profile doesn’t matter, I tested with matching external monitors scale, 8-bits and exact refresh rate,
  • Discrete GPU: NVIDIA Corporation GP107GLM [Quadro P1000 Mobile] (Drives external outputs directly via HDMI/Thunderbolt ports, 1080p, 100% Scale or other, doesn’t matter) 4GB VRAM
  • Memory: 64 GB RAM ECC

Software Profile:

  • OS: Kubuntu 26.04 LTS (Resolute)
  • Display Server: Pure Wayland Session (Plasma 6.6.4 / KWin)
  • Kernel: Linux 7.0.0-22-generic (And verified identically on Linux 6.8.0-111-generic and 7.0.0-15-generic)
  • NVIDIA Driver: Proprietary 580.159.03 (PRIME On-Demand Profile active), also tested other 580 versions with no luck

Symptoms:

  • Booting, login (SDDM), and standard desktop rendering work perfectly fine across multiple monitors.
  • Offloading heavy applications (Photoshop via Wine, local AI models via Ollama) to the Quadro card works flawlessly while keeping the desktop fluid.
  • Moving windows across display boundaries works properly as long as the mouse pointer stays clear of the edge.
  • Moving the mouse pointer close to or over the edge margin results in an instant, unrecoverable freeze of the user session, crashing out (possible revert back to the SDDM greeter screen with keyboard combination, but only to soft restart, as logging in again results in unrecoverable freeze)

Diagnostic Attempts and Proof:

  1. Setting KWIN_FORCE_SW_CURSOR=1 inside /etc/environment completely prevents the crash, but severely throttles the 4K composite performance, making the UI unusable. Hence it is not an acceptable solution, the performance hit is huge.
  2. Setting NoHardwareCursors=true inside [Wayland] section of kwinrc fails to catch the boundary transition, resulting in the same crash.
  3. Enabling KWin’s GPU view update debugging tools reveals no other software hooks to mouse events, it’s clean.
  4. Setting the SDDM login screen to run over an X11 backend allows the mouse to cross displays safely during login screen only, proving the physical hardware links are fine. The freeze triggers purely upon logging back into the KWin Wayland user environment.
  5. Obviously, nvidia_drm.modeset=1 and nvidia_drm.fbdev=1 , tried as modprobe conf and directly in GRUB params, no difference there. Tried multiple other flags and options, no luck. Only X11 or Wayland with KWIN_FORCE_SW_CURSOR=1 work.

Relevant Concurrent Journald Crash Traces:
kwin_scene_opengl: 0x502: GL_INVALID_OPERATION error generated. image and target are incompatible
kwin_scene_opengl: Invalid framebuffer status: “GL_FRAMEBUFFER_INCOMPLETE_ATTACHMENT”
kernel NULL pointer dereference, address: 0000000000000028
RIP: 0010:_nv000582kms+0x4/0x10 [nvidia_modeset]
note: kwin_wayland[5925] exited with irqs disabled
sddm Auth: sddm-helper exited with 255 / coredump in libEGL_mesa.so.0 and libnvidia-egl-wayland.so.1

Please treat this seriously, I spent many days trying to debug this and solve with all I could find over the web. My research was pristine, I didn’t miss anything. Therefore I limited the solutions to X11 session (that’s not an acceptable option for me, I use Wayland since 2022 with success until the 580 drivers, and also Ubuntu is moving away from Wayland) and KWIN_FORCE_SW_CURSOR=1 (also not acceptable as it’s causing everything 10 x slower than X11 session itself, so it’s even a worse “fix”).

Let me say again that on my setup it worked flawlessly under Kubuntu 21.x and Wayland and 470x Nvidia drivers. Due to the Pascal architecture, I cannot use Nvidia OPEN drivers, so my only way to upgrade is wait for Cannonical to release newer version. But my research tells that newer drivers won’t fix that, as apparently Nvidia isn’t focused on this bug at all.

nvidia-bug-report.log.gz (1.5 MB) I assume you will ask for obvious things and the log. To answer your questions/requests: Yes, I turned my computer off and on again :-) And yes, I exported the log, here you are.