https://bugs.kde.org/show_bug.cgi?id=525886
--- Comment #8 from Max Nomad <[email protected]> --- Attached: drm_info output from my machine (drm_info 2.6.0, card0 = nvidia-drm GA106M RTX 3060 Mobile, driver 580.178.04; card1 = amdgpu Cezanne iGPU). Captured in a normal running session. It reflects the current setup, where KWin renders on the AMD iGPU (see below). On the packaging recommendations from comment 5: this machine is on the 580 branch, so the PreserveVideoMemoryAllocations path applies, and it already matches. * nvidia-drm modeset=1 (modprobe.d) * NVreg_PreserveVideoMemoryAllocations=1 (confirmed in /proc/driver/nvidia/params) * NVreg_TemporaryFilePath=/var * nvidia-suspend.service, nvidia-resume.service and nvidia-hibernate.service are all enabled * proprietary kernel module, GSP firmware off (NVreg_EnableGpuFirmware=0) New data that I think settles the driver question. On 2026-10-06 I moved KWin from the NVIDIA dGPU to the AMD iGPU (KWIN_DRM_DEVICES order). Wayland clients follow the compositor, so plasmashell now renders on the AMD render node as well. Nothing else changed: same Plasma 6.6.6, same Qt 6.10.2, same kernel series, same NVIDIA 580.178.04 module loaded. * 2026-09-19 to 2026-10-06, plasmashell on NVIDIA: context lost on 31 of 44 resumes, always 13 to 17 s after "PM: suspend exit". * 2026-10-06 to today, plasmashell on AMD: 0 context losses in 63 resumes. So the bug follows the GPU, not the Plasma/Qt version. I also need to correct my comment 1. I argued there that this looked like a Qt Quick recovery gap, because KWin survived on the same GPU. A later gdb capture of a frozen (non-crashing) instance showed the plasmashell main thread blocked ten frames deep inside libnvidia-eglcore.so.580.178.04, called from QRhi::endFrame() in QSGGuiThreadRenderLoop::renderWindow. That fits a driver fault, and I withdraw that argument. The capture and an nvidia-bug-report log are in my NVIDIA forum report: https://forums.developer.nvidia.com/t/384806 (no NVIDIA response yet). The shutdown crash from comment 1 is now bug NNNNNN, as requested. One related note for anyone on Qt 6.10: some of the crashes when plasmashell is stopped after a context loss are a Qt use-after-free that is already fixed upstream. QSGGuiThreadRenderLoop::handleDeviceLoss() destroyed the QRhi but left QQuickWindowPrivate::rhi dangling, and QSGRenderLoop::postJob() later calls makeThreadLocalNativeContextCurrent() on it. Fix: qtdeclarative 6a2e413a5a "Scenegraph: Nullify QQuickWindowPrivate::rhi on RHI destruction", in Qt 6.11; picked to the 6.10 branch as 6cb7ca6b3d, but not in any 6.10 release. Ubuntu 26.04 ships Qt 6.10.2, so it is affected. -- You are receiving this mail because: You are watching all bug changes.
