Bug#1152416: nvidia-driver: Blackwell shutdown hangs in legacy VBIOS

[vladussie at gmail.com]

Package: nvidia-driver
Version: 615.71.10-1
Severity: normal
Tags: upstream

Dear Debian NVIDIA maintainers,

On Debian 13, a GeForce RTX 5060 Ti in Legacy BIOS/CSM mode hangs during
console restoration when shutting down or unloading nvidia_drm. Preserving
GRUB's graphical framebuffer with GRUB_GFXPAYLOAD_LINUX=keep restored normal
shutdown in a controlled user test. I am reporting this as additional
Debian-specific evidence and a possible documentation/workaround candidate.

Important packaging qualification: 615.71.10-1 is not the stock trixie
NVIDIA driver. I am not claiming reproduction with Debian's stock 550-series
package. Please reassign/close as appropriate if this is outside the scope
of the Debian package; the upstream issue is already known:
https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1222

Environment:
- Debian GNU/Linux 13 (trixie), amd64
- Linux 6.12.111+deb13-amd64, Debian kernel 6.12.111-1
- grub-pc, grub-common, grub2-common: 2.12-9+deb13u2
- systemd: 257.13-1~deb13u1
- NVIDIA open kernel module and core user-space libraries: 615.71.10
- GPU: Palit GeForce RTX 5060 Ti, PCI ID 10de:2d04 (GB206)
- Motherboard: ASUS PRIME H510M-D, BIOS 2803 (2025-12-09)
- Boot mode: Legacy BIOS; /sys/firmware/efi is absent
- Cinnamon/X11, LightDM
- nvidia_drm modeset=1 and fbdev=1

Original behavior:
Normal poweroff stops services and reaches systemd-shutdown. The display
turns off, but the case/fans remain powered for more than three minutes.
Windows shuts down normally on the same machine. The shutdown problem was
also observed with NVIDIA 590.48.01, before upgrading to 615.71.10.

A text-only boot with nvidia/nvidia_drm/nvidia_modeset/nvidia_uvm and
nouveau/nova_core blacklisted powered off normally. This was a diagnostic
control, not a proposed solution.

Direct reproduction while journald and the root filesystem were available:
1. Close the graphical session by stopping LightDM.
2. Stop BOINC and nvidia-persistenced.
3. Verify that no process has open /dev/nvidia* or /dev/dri/* descriptors.
4. Detach the framebuffer vtconsole (write 0 to its bind attribute).
5. rmmod nvidia_uvm succeeds.
6. rmmod nvidia_drm hangs; its process stays in R state.

The kernel watchdog captured the following chain, including a soft lockup
of 208 seconds in the rmmod process (addresses and local identifiers omitted):

  nv_drm_remove_devices.cold [nvidia_drm]
  nv_drm_dev_unload [nvidia_drm]
  ReleaseOwnership [nvidia_modeset]
  nvkms_ioctl_from_kapi [nvidia_modeset]
  nvKmsIoctl [nvidia_modeset]
  ReleaseOwnership [nvidia_modeset]
  RestoreConsole [nvidia_modeset]
  nvFreeCoreChannelEvo [nvidia_modeset]
  nvRmVTSwitch [nvidia_modeset]
  nvRmApiControl [nvidia_modeset]
  nvkms_call_rm [nvidia_modeset]
  Nv04ControlKernel [nvidia]
  _nv04ControlWithSecInfo [nvidia]
  rmapiControlWithSecInfoTls [nvidia]
  rmapiControlWithSecInfo [nvidia]
  _rmapiRmControl [nvidia]
  serverControl [nvidia]
  gpuresControl_IMPL [nvidia]
  resControl_IMPL [nvidia]
  deviceCtrlCmdOsUnixVTSwitch_IMPL [nvidia]
  unixCallVideoBIOS [nvidia]
  nv_vbios_call [nvidia]
  X86EMU_exec [nvidia]
  x86emuOp_in_word_AX_DX [nvidia]
  os_io_read_dword [nvidia]

Before the workaround, the boot log showed:
  Console: colour VGA+ 80x25

Controlled workaround:
A separate GRUB entry used the exact same kernel, initramfs, Linux command
line and NVIDIA module parameters. Only these GRUB commands were added:

  load_video
  set gfxpayload=keep

The user confirmed normal shutdown with this entry. This matches another
user's independently reported workaround in upstream issue #1222.
The boot log then showed a VESA 1920x1080x32 mode before NVIDIA initialized.
There were vesafb reservation/ioremap warnings, so this report does not
claim that vesafb itself successfully registered; the confirmed difference
is the boot display handoff and the observed shutdown result.

The persistent local equivalent now uses:
  /etc/default/grub.d/90-nvidia-framebuffer.cfg:
    GRUB_GFXPAYLOAD_LINUX=keep
  followed by update-grub.

grub-mkconfig correctly emits linux_gfx_mode=keep and load_video/gfxmode for
all Linux entries. No driver options, initramfs changes, forced module
removal or custom shutdown hooks are needed for this local workaround.
The permanent generated configuration has been statically verified; future
updates and all power-management scenarios have not been exhaustively tested.

Please consider documenting the Legacy BIOS/Blackwell interaction and the
per-host GRUB workaround, or coordinating with upstream on a bounded/safe
console-restoration path. This is not a request to force gfxpayload=keep
on all Debian systems. Full raw logs are intentionally not attached to
avoid exposing unrelated host/user information.

Thank you.