building kernel module extensions with correct signing

[devsnek]

Like #227 I'm trying to use NVIDIA's vGPU guest drivers, and I haven't been able to figure out how to get everything to run nicely. I'm mainly confused about how to "build the world" with my pkgs and extensions forks. Are there instructions somewhere for how to do this? Also is there a recommended way to set things up that is simpler than needing whole forks of all these repos?

[smira]

Moved this this to be a discussion.

There are some docs on Talos side:

• https://www.talos.dev/v1.11/advanced/building-images/
• https://www.talos.dev/v1.11/advanced/customizing-the-kernel/
• https://www.talos.dev/v1.11/advanced/kernel-module/

[devsnek]

Thanks for the info, this is really helpful. Unfortunately the build fails when I get to running make installer-base imager PLATFORM=linux/amd64 INSTALLER_ARCH=amd64 REGISTRY=127.0.0.1:5005 PKG_KERNEL=$KERNEL_IMAGE PUSH=true

0.109 install: cannot stat 'usr/lib/modules/6.12.38-talos/kernel/drivers/infiniband/hw/irdma/irdma.ko': No such file or directory
0.219 install: cannot stat 'usr/lib/modules/6.12.38-talos/kernel/drivers/watchdog/f71808e_wdt.ko': No such file or directory

Is it maybe a kernel version mismatch? I checked out v1.11.0 in all 3 repos so I expected them to align.

[devsnek]

ah switching to the release-1.11 branch seems to have fixed it. so the 6.12.38 vs 6.12.43 was the issue. maybe one of the release tags was put on the wrong commit?

[smira]

Our tags are correct, but you should match pkgs with Talos. pkgs are not tagged for each Talos release, see PKGS ?= in Talos Makefile.

[devsnek]

i also had to modify the nvidia-container-toolkit extension to use the 580 driver to match the grid guest drive i'm using, but incredibly enough it seems to be working. i haven't solved nvidia-gridd licensing though, i'll post an update here when i figure out how to get that running (i presume it will mostly involve copying the extension's configuration for nvidia-persistenced)

[devsnek]

As it turns out, running nvidia-gridd in a normal container is enough to get the license installed correctly, so I'll just stick with that.