Display freeze: nvidia-modeset hangs in nvWriteGpEntry from DIFR prefetch after external monitor hotplug

已关闭
  • 哪里不对？
#1,503 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
48/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
c, linux, ubuntu
领域
devtools, operating-systems

调研方向

Start by reviewing the reported stack in nvWriteGpEntry, PrefetchHelperSurfaceEvo, and nvDIFRPrefetchSurfaces, then examine the attached kernel logs for the hotplug sequence. The report describes the failure path and reproduction steps, but does not identify source files or tests; reproduce across monitor disconnect and reconnect and confirm modesetting no longer hangs.

由索引模型根据 Issue 内容生成。

描述

bug
NVIDIA Open GPU Kernel Modules Version

615.71.09 (also reproduced on 595.99.02), Ubuntu packages nvidia-driver-615-open / nvidia-driver-595-open

Please confirm this issue does not happen with the proprietary driver (of the same version). This issue tracker is only for bugs specific to the open kernel driver.
  • I confirm that this does not happen with the proprietary driver package.
Operating System and Version

Ubuntu 26.04

Kernel Release

7.0.0-38-generic

Please confirm you are running a stable release kernel (e.g. not a -rc). We do not accept bug reports for unreleased kernels.
  • I am running on a stable kernel release.
Hardware: GPU

GeForce RTX 4090 Laptop GPU (AD103M, 10de:2757), hybrid graphics with Intel Raptor Lake-S UHD (i915). Alienware m18 R2, BIOS 1.23.0.

Describe the bug

The whole desktop locks up after the external monitor (Samsung U28E590 on HDMI-A-1, which is wired to the NVIDIA GPU) disconnects and reconnects. The monitor reconnect coincides with a USB UGREEN KVM hub reconnecting. The kernel stays alive (SysRq works), but the display is frozen and switching to a VT doesn't work either. The only recovery is REISUB or a hard power-off.

Desktop: GNOME Shell 50.1 on Wayland, with nvidia-drm.modeset=1.

The kernel's hung-task detector shows mutter's KMS thread stuck in nv_drm_atomic_apply_modeset_config, waiting on a semaphore held by the nvidia-modeset kthread. That kthread is runnable (state R) and spins inside the DIFR prefetch path. Every snapshot (six, taken 20 seconds apart, plus SysRq+W) shows the same stack:

RIP: 0010:nvWriteGpEntry+0x.../0x370 [nvidia_modeset]
 nvPushKickoff+0x28/0x50 [nvidia_modeset]
 PrefetchHelperSurfaceEvo+0x45c/0x650 [nvidia_modeset]
 nvDIFRPrefetchSurfaces+0x12f/0x1f0 [nvidia_modeset]
 DifrPrefetchEventDeferredWork+0x16/0x30 [nvidia_modeset]
 nvkms_kthread_q_callback+0xe9/0x190 [nvidia_modeset]
 _main_loop+0x99/0x160 [nvidia_modeset]

It looks like nvWriteGpEntry is waiting for GPFIFO space that the GPU never frees after the hotplug, and it holds the nvkms lock while it spins, so every modeset from the compositor blocks.

Timeline of the 615.71.09 occurrence (2026-10-29, local time):

  • 15:17:54 external monitor and USB hub disconnect
  • 15:30:07 monitor and hub reconnect
  • ~15:31:27 display frozen; hung-task reports start at 15:31:47
  • 15:33:25 SysRq+W captured, then REISUB

There are no Xid errors in the log. There's no module parameter or registry key that I can find to turn DIFR off as a workaround.

To Reproduce
  1. Laptop with hybrid graphics, external monitor connected to an output driven by the NVIDIA GPU, GNOME Wayland session.
  2. Disconnect the external monitor (or let it drop off) and reconnect it.
  3. It doesn't happen every time. Several reconnects succeed, but it has happened 4 times in about 4 days.
Bug Incidence

Sometimes

nvidia-bug-report.log.gz

nvidia-bug-report-redacted.log.gz

Attached, with hardware serials, MAC addresses, the monitor EDID and user/host names redacted. The report was generated after the reboot, so it doesn't contain the hang itself. The kernel log from two hangs (on 595.99.02 and 615.71.09) is attached separately as nvidia-freeze-kernel-log-redacted.log.gz.

nvidia-freeze-kernel-log-redacted.log.gz

More Info

Happy to test patches or collect more data (for example with nvidia_modeset debug=1).

主要语言
C
星标
17.5k
派生
1.9k
30 天内没有已合并 PR

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

NVIDIA/open-gpu-kernel-modules 的其他 Issue

查看 NVIDIA/open-gpu-kernel-modules 的全部 Issue

相似的 Issue

更多 C Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。