Downgrading nvidia driver from 535

Ask Fedora
, , ,
1

I’m trying to downgrade my nvidia driver from 535, since it is causing regular stuttering and frame drop issues that weren’t there before.

Steps I’ve taken:

  • Remove current driver: sudo dnf remove \*nvidia\* --exclude nvidia-gpu-firmware, then reboot
  • Lock out the problematic version: sudo dnf install 'dnf-command(versionlock)', then sudo dnf versionlock exclude akmod-nvidia-535.113.01-1.fc38
  • Install drivers following the normal instructions and verifying that all packages are nvidia 350.x and versionlocking anything that tries to install a different version as above
  • Reboot

I was surprised that despite everything apparently working, I’m still on nouveau after a reboot. lsmod | grep nouveau shows it loaded, and lsmod | grep nvidia shows nothing. All expected packages are installed:

 $ dnf list --installed | grep nvidia
akmod-nvidia.x86_64                                       3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
libva-nvidia-driver.x86_64                                0.0.10-3.fc38                          @updates                                         
nvidia-gpu-firmware.noarch                                20230919-1.fc38                        @updates                                         
nvidia-persistenced.x86_64                                3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
nvidia-settings.x86_64                                    3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia.x86_64                                3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia-cuda.x86_64                           3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia-cuda-libs.i686                        3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia-cuda-libs.x86_64                      3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia-kmodsrc.x86_64                        3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia-libs.i686                             3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia-libs.x86_64                           3:530.41.03-1.fc38                     @rpmfusion-nonfree                               
xorg-x11-drv-nvidia-power.x86_64                          3:530.41.03-1.fc38                     @rpmfusion-nonfree

So I tried manually triggering the akmod build, but it fails:

 $ sudo akmods
Checking kmods exist for 6.5.8-200.fc38.x86_64             [  OK  ]
Building and installing nvidia-kmod                        [FAILED]
Building rpms failed; see /var/cache/akmods/nvidia/530.41.03-1-for-6.5.8-200.fc38.x86_64.failed.log for details

Hint: Some kmods were ignored or failed to build or install.
You can try to rebuild and install them by by calling
'/usr/sbin/akmods --force' as root.

The log file for the failed build is here, the relevant error seems to be a lot of variations on this error:

2023/10/29 18:57:13 akmodsbuild: /tmp/akmodsbuild.M8wOQ9Bg/BUILD/nvidia-kmod-530.41.03/_kmod_build_6.5.8-200.fc38.x86_64/common/inc/nv-mm.h:88:16: error: too many arguments to function 'get_user_pages'
2023/10/29 18:57:13 akmodsbuild:    88 |         return get_user_pages(current, current->mm, start, nr_pages, write,
2023/10/29 18:57:13 akmodsbuild:       |                ^~~~~~~~~~~~~~
2023/10/29 18:57:13 akmodsbuild: ./include/linux/mm.h:2430:6: note: declared here
2023/10/29 18:57:13 akmodsbuild:  2430 | long get_user_pages(unsigned long start, unsigned long nr_pages,
2023/10/29 18:57:13 akmodsbuild:       |      ^~~~~~~~~~~~~~

My reading of this is that the akmod for version 530.41.03 is failing to build against my kernel version (since the function in the error seems to be from a kernel header).

Can anybody help me finish this downgrade? Is it just not possible to downgrade without also downgrading my kernel, and if so, how can I figure out which kernel version to revert to?

System info:

  • Fedora 38 Workstation, GNOME, X11
  • Kernel: 6.5.8-200.fc38.x86_64 #1 SMP PREEMPT_DYNAMIC Fri Oct 20 15:53:48 UTC 2023 x86_64 GNU/Linux
  • GPU: Nvidia 3070Ti (desktop)
2

You will need to also use an old kernel I would guess.
Or you will need to patch the nvidia glue code to use the changed way to call get_user_pages.
You could look at the latest driver for what is required.

3

What’s the best way of me finding the source of what’s being compiled?

Any idea how I might find which kernel version each is targeted for?

4

The name of the .c file will be in the akmodsbuild log.

From your pastbin I see this:

2023/10/29 18:57:13 akmodsbuild: In file included from /tmp/akmodsbuild.M8wOQ9Bg/BUILD/nvidia-kmod-530.41.03/_kmod_build_6.5.8-200.fc38.x86_64/common/inc/nv-pgprot.h:30,
2023/10/29 18:57:13 akmodsbuild:                  from /tmp/akmodsbuild.M8wOQ9Bg/BUILD/nvidia-kmod-530.41.03/_kmod_build_6.5.8-200.fc38.x86_64/common/inc/nv-linux.h:33,
2023/10/29 18:57:13 akmodsbuild:                  from /tmp/akmodsbuild.M8wOQ9Bg/BUILD/nvidia-kmod-530.41.03/_kmod_build_6.5.8-200.fc38.x86_64/nvidia/nv-nano-timer.c:31:
2023/10/29 18:57:13 akmodsbuild: ./include/linux/mm.h:2430:35: note: expected 'long unsigned int' but argument is of type 'struct task_struct *'
2023/10/29 18:57:13 akmodsbuild:  2430 | long get_user_pages(unsigned long start, unsigned long nr_pages,
2023/10/29 18:57:13 akmodsbuild:       |                     ~~~~~~~~~~~~~~^~~~~

So it should be in nv-nano-timer.c:31

5

This shows the system did not build and install the kmod-nvidia package for the currently installed nvidia packages (530.41.03).

Try that as sudo akmods --force and if it still fails then try
sudo akmods --kernels $(uname -r) --force

The same suggestion was at the bottom of what you posted for the failed akmods run.

6

I should have said, I tried akmods --force and it failed with exactly the same compilation error. Wouldn’t the default kernel for akmods be the current one anyway?

7

I meant where can I find the repo that contains the source code, not which file within the code is failing.

8

On my system I find

# find /usr -name nvidia-kmod*
/usr/share/nvidia-kmod-535.113.01
/usr/share/nvidia-kmod-535.113.01/nvidia-kmod-535.113.01-x86_64.tar.xz
/usr/src/akmods/nvidia-kmod.latest
/usr/src/akmods/nvidia-kmod-535.113.01-1.fc39.src.rpm

and 

ls -l /usr/src/akmods/
total 240
-rw-r--r--. 1 root root 90454 Sep 21 19:00  nvidia-kmod-535.113.01-1.fc39.src.rpm
lrwxrwxrwx. 1 root root    37 Sep 21 19:00  nvidia-kmod.latest -> nvidia-kmod-535.113.01-1.fc39.src.rpm

Those files come from installing akmod-nvidia from rpmfusion.

# dnf provides /usr/src/akmods/nvidia-kmod-535.113.01-1.fc39.src.rpm
Last metadata expiration check: 0:08:46 ago on Mon 30 Oct 2023 01:16:52 PM CDT.
akmod-nvidia-3:535.113.01-1.fc39.x86_64 : Akmod package for nvidia kernel module(s)
Repo        : @System
Matched from:
Filename    : /usr/src/akmods/nvidia-kmod-535.113.01-1.fc39.src.rpm

akmod-nvidia-3:535.113.01-1.fc39.x86_64 : Akmod package for nvidia kernel module(s)
Repo        : rpmfusion-nonfree
Matched from:
Filename    : /usr/src/akmods/nvidia-kmod-535.113.01-1.fc39.src.rpm

akmod-nvidia-3:535.113.01-1.fc39.x86_64 : Akmod package for nvidia kernel module(s)
Repo        : rpmfusion-nonfree-nvidia-driver
Matched from:
Filename    : /usr/src/akmods/nvidia-kmod-535.113.01-1.fc39.src.rpm