Bug#802246: Can't modprobe -r nvidia, rmmod works

[Luca Boccassi]

On 18 October 2015 at 22:18, Andreas Beckmann <anbe at debian.org> wrote:
> On 2015-10-18 23:05, Luca Boccassi wrote:
>> I installed 340.93-5 on my laptop with optimus. Any other solution you
>> would like me to try out?
>>
>> Another solution would be scrapping the bumblebee patch we ship to use
>> modprobe instead of rmmod, but it would require fixing the code to
>> remove the nvidia-uvm (and in the future nvidia-modeset too),
>> otherwise it would break for Cuda users, which is the reason we
>> introduced the patch in the first place.
>
> glx-alternatives 0.6.92 from git
>
> this comes with a lower priority nvidia/bumblebee alternative for glx
> which needs to be activated manually
>
>
> always run
>   dpkg-reconfigure glx-alternative-nvidia
> after
>   update-alternatives --config glx
>
> (and similarly reconfigure nvidia-alternative (or
> nvidia-legacy-*-alternative) after changing the nvidia alternative)
>
> to activate and run the triggers
>
> (THIS NEEDS TO BE DOCUMENTED)

This works with 340.93-5. Also tested on the desktop for collateral
and it seems fine.

Would this mean that a bumblebee user has to run this manually when
the package is installed? Is there a way we can automate this and make
the bumblebee package trigger it?

> since that will 'modprobe -r nvidia'/'modprobe nvidia' you may end up
> with the wrong^Wproblematic permissions - are you in the video group?

The devices are created with 666:

$ ls -l /dev/nvidia*
crw-rw-rw- 1 root root 195,   0 Oct 18 23:17 /dev/nvidia0
crw-rw-rw- 1 root root 195, 255 Oct 18 23:17 /dev/nvidiactl
crw-rw-rw- 1 root root 195, 254 Oct 18 23:15 /dev/nvidia-modeset

Also, strangely, the uvm (and modeset when testing 358) device remains
even after the modules are unloaded and the card is turned off:

$ ls -l /dev/nvidia*
crw-rw-rw- 1 root root 195, 254 Oct 18 23:15 /dev/nvidia-modeset
crw-rw-rw- 1 root root 244,   0 Oct 18 23:20 /dev/nvidia-uvm

nvidia0 and nvidiactl correctly disappear. It doesn't seem to cause
any issue, I can repeatedly run tests for both Cuda and OpenGL
programs and the kernel modules are always unloaded and the card is
always turned off. Bit strange though.

Kind regards,
Luca Boccassi