Shutdown/reboot issue with nvidia-open-dkms and RTX 5070ti

advesperascit

I recently upgraded from a GTX 1070 to a RTX 5070ti, and as a result also switched from the nvidia-580xx-dkms drivers to nvidia-open-dkms. After this change the computer no longer completely shuts down or completes a reboot, it remains powered on with the screens turned off. I have to hold down the power button to turn it off completely.

I've narrowed it down to the nvidia drivers: if I log off, switch TTY and stop SDDM, I can run modprobe -r nvidia nvidia_drm nvidia_modeset nvidia_uvm which kills all video output, then ctrl-alt-delete will trigger a successfully reboot that does not hang.

Is there a way to add this modprobe command for s6 to run during shutdown?

eric

[Main]
Type = oneshot
RequiredBy = ( sddm )
TimeoutStop = 5000
[Start]
Execute = ( true )
[Stop]
Execute=( modprobe -r nvidia nvidia_drm nvidia_modeset nvidia_uvm )

At start time (a no-op here) it will be launched before sddm, at stop time it will be stopped after sddm.

Anyway, this is an hack, it doesn't solve the bug from Nvidia. This is not depending of the init system(you will have the same trouble with systemd). Anything else holding /dev/nvidia* will make the removal fail with EBUSY and you're back to the hang. If you run CUDA services, nvidia-persistenced, ollama, etc., they need to be in that RequiredBy too.

advesperascit

Just a heads up for anyone experiencing the same issue: it should be modprobe -r nvidia_drm nvidia_modeset nvidia_uvm nvidia (nvidia last, or there'll be an error), I was careless and wrote it wrong in the OP.

eric This is not depending of the init system(you will have the same trouble with systemd)

Indeed; I got the modprobe solution from a forum post of a systemd user experiencing the same issue, remedied through a systemd shutdown script.

Thank you for the help, unfortunately it doesn't work as I'd hoped. It does unload the drivers as the screens go black and turn off, but the shutdown/reboot does not complete. I do get the same issue when I try # modprobe -r nvidia_drm nvidia_modeset nvidia_uvm nvidia ; poweroff instead of using ctrl-alt-delete. I'll troubleshoot some more when I get time (maybe modprobe -rf will fare better), unsurprisingly nvidia-open has the same issue as it's dkms sibling.