Bug#960163: nvidia-driver: x86/modules: Skipping invalid relocation

[Andreas Beckmann]

Control: severity -1 normal
Control: tag -1 unreproducible

On 11/05/2020 05.24, Allan Wind wrote:
> The good news is that I have the nvidia module working again,
> and it's probably a weird edge case:
> 
> (1) nvidia-drivers version 418.113-1 is not compatible with linux   
> 4.19.0-5 headers and fail to load with relocation error, and

That's weird, but an completely untested combination and not worth
further investigation.

> (2) When I was running 4.19.0-9-amd64 (with nouveau),
>    `dpkg-reconfigure nvidia-kernel-dkms` rebuild the nvidia    module
> against the -5 headers and they ended up in   
> /lib/modules/4.19.0-9-amd64 which failed to load (i.e. (1)).

That should not happen ... and I couldn't reproduce something like that
in a buster chroot.

> 
> Since my attempt yesterday, I purged the -5 kernel, and
> `dpkg-reconfigure nvidia-kernel-dkms` now failed with missing
> headers.  I then purged the -5 headers and installed the -9 headers and
> upon reboot nvidia module was loaded.
> 
> I purged linux-image-amd64 a while ago when I explored kernel/nvidia
> combinations as possible work-around for another bug #905309 (and
> #922497).  This meant that I didn't get -9 and hit (1) when I thought I
> was doing a low-risk (stable) upgrade.

It's recommended to install linux-image-amd64 and linux-headers-amd64 to
track the current versions on upgrades.

> Re (1) I don't know _if_ or _how_ Debian should address this
> actually, it's incompatibility with the combination of nvidia-driver
> 418.113-1 and linux-headers-4.19.0-5 that only shows up at run-time on
> the subsequent reboot.

I won't spend time on this unless it happens with current versions, too.

> Re (2) seems like a straight build defect.  If linux-image version x is
> installed then linux-headers x is used to build a module
> for linux-image y (probably as a fallback).  This fails
> with a pre-build check if linux-image x not installed.

I tried something with 4.19.0-9-amd64 and 4.19.0-8-amd64 in a chroot,
but couldn't get modules built for a kernel where no headers were available.


Andreas