Gigabyte Aero 15x v8 Notes and Scripts

I have gotten this laptop to reliably run Linux with some workarounds and caveats. I am using Fedora 28 / Kernel version 4.18 with XWayland.

The laptop has crappy ACPI support for Linux, including issues with powering down the dGPU and issues with the keyboard Fn keys. Other than that, it seems to work pretty well.

dGPU / Optimus support

This laptop is plauged by ACPI power issues that result in hard kernel lockups without some workarounds. This is apparently a common problem with NVIDIA Optimus laptops that has persisted over the past 2 years.

For details on this problem, this bug report for nouveau is the best source of information.

This issue prevents bbswitch / bumblebee from working. When the card is off, the kernel locks up whenever any program tries to initiate communication with the card (i.e. lspci crashes the system.) There is a github issue that details the problem; it's the same issue as the referenced bug in nouveau. None of the acpi_osi workarounds were successful on the Aero 15x v8, presumably because it doesn't support Windows 7.

Nouveau also causes kernel lockups when suspending / resuming the system. Karol Herbst has provided a patch that fixes this issue with Nouveau in my setup. Nouveau will power off the card when it is not in use.

What works on my setup:

• Attaching dGPU to primary display with NVIDIA proprietary drivers using the modesetting driver (when nvidia module is loaded at boot time).
• Attaching dGPU to primary display using noveau (when noveau is loaded at boot time).
• Powering on/off the card works by dynamically loading / unloading the nvidia and nouveau modules while XWayland is bound to the Intel card on the primary display. This works for CUDA/OpenCL workloads and the NVIDIA Docker runtime.

What doesn't work:

• X11 gets angry about dynamically loading and unloading nvidia and nouveau, so this only works with XWayland. As a result, nvidia-xrun doesn't work.
• bumblebee doesn't work because of the dependency on bbswitch.

Caveats:

• Booting with either nouveau or nvidia loaded makes it hard to unload the modules later, even if the user logs out and logs back in. To get around this, I blacklist the modules at boot time if I want to use the Intel graphics for the primary display. We load noveau using GDM auto-start instead of at boot time, preventing XWayland from grabbing it and using it to render the screen, so we can display things with Intel and nouveau can turn the card off. This requires an edit to /etc/sudoers that allows wheel members to load nouveau without a password. Something like:

%wheel	ALL=(ALL)	NOPASSWD: /usr/sbin/modprobe nouveau

• On the flipside of this issue, in order to use the nvidia module to drive the primary display, the modules must be loaded at boot time. In order to disable the card after starting up this way, a reboot is required.

Helper scripts:

• build-nouveau.sh: This script will download the Linux kernel from git.kernel.org, patch the nouveau module with Karol's power management workaround, build it against your running kernel, and install the new module.
• gpuswitch.sh: Loads and unloads modules and shuffles the config files around to turn the dGPU on or off.
• gpuswitch-configs: Config files that gpuswitch.sh uses, please read it to figure out where they go.

Since the NVIDIA docker runtime works so well on this system, I was going to attempt to pass a TTY to a Docker container and try an nvidia-xrun-style setup with X11 inside a container.

Keyboard support

The keyboard sends out sensible HID codes for the Fn volume controls and Sleep button, but the scan codes for the other buttons do not follow the HID standard as I understand it. They (mostly) do not fire events on key release and the format of the scan codes doesn't conform to the HID documentation. The usbhid driver doesn't know what to do with the messages and drops them on the floor; we don't even see "unknown key" messages in dmesg. We can only see them if we record the raw messages from the kernel using /dev/hidraw.

What's even weirder is that the Volume keys fire normal HID events AND these non-standard events. They are also the only keys that fire a key press and key release in the non-standard format. It's possible these are WMI events, but the keys don't do anything in Windows either. I don't know much about WMI and I am exploring this further; if possible I would like to add support in the kernel.

The only keys that work in Windows are the Volume keys, the Sleep key, and the Fn + F5 key which brings up a display control panel. Fn + F5 doesen't show up on /dev/hidraw in linux, though I can see the keyboard is sending messages over the USB bus.

The only vendor-supported way to get these keys working in Windows is to run the Gigabyte "Smart Manager" software. It's buggy and terrible. I unpacked the exe and it comes with it's own DLLs... how fun! I may reverse engineer this software further.

The other functionality that is missing is the LED control. I think the most direct way to go about this would be to hook up the control software to a debugger.

Until the driver is working, I have done some hacky stuff with hid-record, awk and bash to get Fn-key support working for my use cases. I've also documented the nonstandard scan codes and mapped them to their keys. See hotkeys.sh for details.

What works:

• Normal keys and numpad
• Break, Insert, PrtScr
• Volume Control Fn keys (F7 - F9)
• Sleep Fn key (F1)
• Sniffing /dev/hidraw and doing hacky things for the other hotkeys

What doesn't work:

• Natve support in usbhid driver for most of the Fn hotkeys.
• Controlling the lights.