gpu-manager fails to detect hybrid requirements of Oryx Pro

[Chris Halse Rogers]

gpu-manager fails to detect that the NVIDIA card of my System76 Oryx Pro requires offloading, resulting in X failing to start when hybrid graphics is enabled in the firmware.

Using the attached xorg.conf, X correctly starts with the NVIDIA card rendering and the Intel GPU doing the eDP output.

ProblemType: Bug
DistroRelease: Ubuntu 17.10
Package: ubuntu-drivers-common 1:0.4.23
ProcVersionSignature: Ubuntu 4.11.0-10.15-generic 4.11.8
Uname: Linux 4.11.0-10-generic x86_64
NonfreeKernelModules: zfs zunicode zavl zcommon znvpair
ApportVersion: 2.20.5-0ubuntu5
Architecture: amd64
CurrentDesktop: GNOME
Date: Fri Jul 21 11:40:43 2017
SourcePackage: ubuntu-drivers-common
UpgradeStatus: No upgrade log present (probably fresh install)

[Chris Halse Rogers (raof)]

This bug was submitted with the nvidia-384 drivers from the driver-updates PPA installed, but the same behaviour is apparent with nvidia-375 from the Artful repositories.

[Alberto Milone (albertomilone)]

Hi Chris, can you reproduce the problem and attach your /var/log/gpu-manager.log, please?

[Chris Halse Rogers (raof)]

Hah!

Unfortunately, I cannot reproduce it at the moment. But I did some digging when I *could* reproduce it:

The problem appears when the nvidia driver binds to /dev/fb0 (which I *think* is caused by the firmware selecting it as the boot VGA device?).

When nvidia is driving /dev/fb0, nothing from the is shown on the display (obviously, as the display is hooked up to the Intel), and gpu-detect does *not* detect that it needs offloading.

By switching the fbcon onto /dev/fb1 (via the fbcon=map:1 kernel parameter) I get VT content and can start X on the Intel card by manually modprobing bbswitch, unloading nvidia, “echo OFF | sudo tee /proc/.../bbswitch”, and then starting GDM.

In this situation, by using the xorg.conf I attached, PRIME offload nvidia/intel works fine.

At the moment i915 is binding to /dev/fb0, and gpu-manager/nvidia-settings is working fine switching between Intel and NVIDIA.

What happens to bind to /dev/fb0 seems to be semi-stable - it tends to persist across reboots and power cycling, but it has flipped a couple of times in the past for some reason. The kernel booed does not appear to be the determining factor.

Next time it boots with NVIDIA binding to /dev/fb0 I'll get you the requested log :)