SOLVED: nvidia-current-dkms and "RT" kernel

[kent_dorfman766]

There are some rumblings about this when doing a web search, but nothing substantive jumps out.

I need to switch to linux-image-rt/oldstable and it seems nvidia-current-dkms is incompatible. Any elegant path forward?

I'd hate to have to jump to manual driver build from the nvidia vendor website.

[sunrat]

Nvidia has never built on RT kernels afaik. There's a check in the build script which blocks it. You can override it with:

Code: Select all

export IGNORE_PREEMPT_RT_PRESENCE=1

before installing a kernel or nvidia driver.

I'm interested to hear why you want to install an RT kernel. They may be useful in an industrial engineering setting but can be a problem for general purpose use. I tried several for audio production a while back but got better results with a Liquorix kernel which is default configured for low latency. Even stock kernels since 5.19 have been capable of full pre-emption by adding the kernel boot option preempt=full.

[kent_dorfman766]

IMHO, the linux kernel has never handled IO channel bottlenecks well, and it seems to have gotten worse the more they try to genericise a kernel "for general use" by playing with newer more complicated schedulers. RT is an experiment to see if the occasional "mouse freeze" or system lag during disk IO goes away or gets better.

Also, my simulation projects are easier to metric if I can measure and mimimize kernel latency in a more deterministic way.

Back when I genned my own kernels 4.x and usually build nvidia drivers from vendor source, I did use rt patches with nvidia drivers but yes, there seems to be some online discussion about nvidia not playing well in that environment.

Will try to hack the dkms build using the env var you mention and see what happens.

Thanks!

[kent_dorfman766]

FWIW, the symbol export previously suggested just changed the error message to:

Code: Select all

root@files:/usr/src# export IGNORE_PREEMPT_RT_PRESENCE=1
root@files:/usr/src# dkms build nvidia-current/470.223.02 -k 5.10.0-28-rt-amd64/x86_64 --kernelsourcedir /usr/src/linux-headers-5.10.0-28-rt-amd64
Error!  The /var/lib/dkms/nvidia-current/470.223.02/5.10.0-28-rt-amd64/x86_64/dkms.conf for module nvidia-current includes a BUILD_EXCLUSIVE directive which
does not match this kernel/arch.  This indicates that it should not be built.

and I'm not inclined to start hacking the dkms module build scripts, so I guess it's a dead issue unless I get really bored.

An interesting anomoly is that the referenced /var/lib/dkms/... directory doesn't even exist. I wonder if the debian installation script was "lacking by design" since they don't want to install under RT anyways.

Thx!

[stevepusser]

Liquorix is fine for low latency I/O for demanding audio/video work; have you at least tried it?

The latest Liquorix kernels are not being built for Bullseye/oldstable since the developer switched to using zstd kernel and module compression, but I continue to do so for MX Linux by reverting to xz for that.

[sunrat]

FWIW, the symbol export previously suggested just changed the error message to:

Code: Select all

root@files:/usr/src# export IGNORE_PREEMPT_RT_PRESENCE=1
root@files:/usr/src# dkms build nvidia-current/470.223.02 -k 5.10.0-28-rt-amd64/x86_64 --kernelsourcedir /usr/src/linux-headers-5.10.0-28-rt-amd64
Error!  The /var/lib/dkms/nvidia-current/470.223.02/5.10.0-28-rt-amd64/x86_64/dkms.conf for module nvidia-current includes a BUILD_EXCLUSIVE directive which
does not match this kernel/arch.  This indicates that it should not be built.

I never tried building it from source. It worked last time (which was when Bullseye was Testing) just by setting the environment variable and apt install nvidia-driver. nvidia-tesla-470-driver is in Debian repo so no need to build from source. I read somewhere they plan to drop the "tesla" moniker soon which is good as it's confusing.

Also +1 what Stevo said about Liquorix, they are well worth trying for improved latency and some other performance tweaks. Only kernel I have used for years.

[kent_dorfman766]

Liquorix is fine for low latency I/O for demanding audio/video work; have you at least tried it?

I've not played with it, but intend to review it now that I'm aware.