Arch Linux User Repository

For testing users, contributors: check out https://github.com/archlinux-jerry/nvidia-340xx
Warning: force push may happen to match the aur tree

@johnstef 5.11 would have the same issues as 5.12. If you want to go back to an old kernel, go to 5.10 using linux-lts and point your bootloader at that kernel. However, I'll bet you just have a bad X11 config like I did.

I can't tell- are you using Wayland and getting these errors via an Xwayland session? Or just regular X11? I only ask because it seems Wayland and nvidia will never get along.

I had essentially identical errors as you, and I was able to fix it by correcting a bad xorg.conf configuration. The nvidia ModulePath needs to be above the default xorg one. Something like:

Section "Files"
  ModulePath   "/usr/lib64/nvidia/xorg"
  ModulePath   "/usr/lib64/xorg/modules"
EndSection

I can't say for sure whether or not you should accomplish this by directly modifying /etc/X11/xorg.conf (like I do) or by placing an "overlay" config file (like the /usr/share/nvidia-340xx/20-nvidia.conf provided in this package) into the /etc/X11/xorg.conf.d directory. It depends on your existing configuration and how any existing ModulePath entries may be stepping on each other.

The goal is for X11 to have nvidia-specific hardware acceleration libraries (libEGL, libGL, etc.) available at runtime otherwise you'll get those errors.

(It is probably an upstream bug that allows a double free to occur when a library is missing instead of gracefully failing, but I digress).

I installed kernel 5.12 and I get the message that EGL could not be initialized on mpv and other apps.

[vo/gpu] Probing for best GPU context.
[vo/gpu/opengl] Initializing GPU context 'wayland'
[vo/gpu/opengl] Initializing GPU context 'x11egl'
[vo/gpu/x11] X11 opening display: :0
[vo/gpu/x11] Display 0 (HDMI-0): [0, 0, 1920, 1080] @ 60.000000 FPS
[vo/gpu/x11] Current display FPS: 60.000000
[vo/gpu/opengl] Could not initialize EGL.
free(): double free detected in tcache 2

I tried downgrading back to 5.11 but that error still exist! Do I have to do anything special to go back to the old kernel?

It works perfect, I had kernel 5.11 with driver 340.108-18 I did not have to modify anything.

OK, I honestly forgot that I included an example config file. Since I don't have the hardware for testing, and since I still don't have a sense from users that it is a requirement, I will leave it in the post_install and depend on the user's ability to read pacman's output and manually copy the example over.

@graysky I noted the advice that was given after installing the module on kernel 5.11: "-> You must tell Xorg to use the nvidia driver with kernels >=5.11.0. -> Minimal xorg config example: /usr/share/nvidia-340xx/20-nvidia.conf -> To place in /etc/X11/xorg.conf.d/ " In fact, as @jayache80 pointed out, it is in /etc/X11/xorg.conf (and not in xorg.conf.d/) that the "Files" section must be filled in. However, I am not able to conclude whether 20-nvidia.conf is mandatory or not. I don't remember to create it myself... Thanks to you. Sincerely.

Translated with www.DeepL.com/Translator (free version)

Not specifically.

I have not been following this closely. Is 20-nvidia.conf a requirement for all setups of this driver with 5.12? Should the package provide an equivalent?

@jayache80 Thank you so much! ...I just added to my xorg.conf the two lines you mention, and archlinux with the latest kernel (uname -mr = 5.12.2-arch1 x86_64) boots well on lightdm and gdm for xfce, gnome and kde. Only sddm doesn't want to show up... Sincerely

I'm an idiot. The order of ModulePath entries was somehow incorrect in my xorg.conf:

As in the provided 20-nvidia.conf, and is nicely called out to the user upon installing this driver, it absolutely needs to be:

Section "Files"
  ModulePath   "/usr/lib64/nvidia/xorg"
  ModulePath   "/usr/lib64/xorg/modules"
EndSection

However, somehow I had it reversed. The following is WRONG:

Section "Files"
  ModulePath   "/usr/lib64/xorg/modules"
  ModulePath   "/usr/lib64/nvidia/xorg"
EndSection

Sorry for all the noise.

The following can be ignored as it was totally a red herring. DRM is just different now and you can expect that it will be initialized differently and you won't see the same system log output in 5.11+ as 5.10.

I think I know why I'm having problems on 5.11/5.12 but not 5.10. I noticed in the system log on 5.10 that DRM is getting initialized properly:

[drm] Initialized nvidia-drm 0.0.0 20150116 for 0000:02:00.0 on minor 0

However, no such output in 5.11/5.12 when using this kernel module. That stuff is taken care of in nv-drm.c starting essentially with the call to nv_drm_pci_init from nv_drm_init but only #if defined(NV_DRM_AVAILABLE). I proved it wasn't being called with some printk's.

Something seems a little strange with the build system as that is defined in 5.10 but is not defined when building 5.11/5.12. I even added

#define NV_DRM_AVAILABLE

to the top of nv-drm.c to no avail. It's probably something really simple and I'll bet this will fix my issues (but I don't know about anyone else's).

Anyone know what's up with that define?

Back in March on 5.11, people were having luck with GTK based greeters but Qt based ones weren't working (e.g. SDDM). So Googling around the error that @jayache80 reported, I found some references to people having weird problems with Qt that were fixed with setting QT_X11_NO_MITSHM=1 in someplace like /etc/environment. Maybe this explains why some people have luck and some don't? I'll admit that this is far-fetched, but it might be worth a shot given that it is low effort to try. The rabbit hole I went down are in the links below. Disabling the nvidia splash screen with Option "NoLogo" "1" in xorg.conf might not help, but it wouldn't hurt either.

https://forum.mxlinux.org/viewtopic.php?t=53653

https://forum.mxlinux.org/viewtopic.php?f=108&t=53603&start=10

https://ubuntuforums.org/showthread.php?t=2423870