[Feature Request] Block GPU, but with allowlists

[BreiteSeite]

Hi,

i have a Gaming PC that doubles as a 24/7 NAS. I would like to have the following if possible:

By default, everything should only be allowed to access my dGPU to make sure no game renders on my iGPU.

However, i also run a Plex Media Server which uses the iGPU from my intel processor for transcoding. This is basically the only process on my system, that should access the iGPU.

So ideally, i'd to represent a config as the following: Every process can only see/access one GPU. For all processes not in the allowlist for the iGPU (Plex, Jellyfin, etc), it would be the dGPU. For processes in the allowlist, it would be only the iGPU.

However, the limiting to Plex only seeing the iGPU is optional. It would be also sufficient to say "block access to the iGPU, but processes in the allowlist basically see the same system like there would be no "block" at all (i.e. hybrid).

$ cardwire gpu 1 --lsof
  /dev/dri/card1: []
  /dev/dri/renderD128: ["Plex Transcoder"]

[luytan]

You can use the manual for that, here's an example on my PC:

> cardwire set manual
Mode has been set to Manual

> cardwire list
ID  NAME                                PCI           RENDER      CARD   DEFAULT DISCRETE  BLOCKED
--  ----------------------------------  ------------  ----------  -----  -------  -------  -------
0   AMD Radeon RX 9070 XT               0000:03:00.0  renderD128  card1  (*)      (*)      false  
1   AMD Ryzen 7 7700X 8-Core Processor  0000:19:00.0  renderD129  card0  ( )      ( )      false   

# Block your iGPU
> cardwire gpu 1 --block
GPU 1 has been blocked

# Here i'm using nvtop to list available GPUs
> nvtop -s |grep "device_name"
   "device_name": "AMD Radeon RX 9070 XT",

# And now i use this env to force the GPU 1
> CARDWIRE_FORCE_GPU=1 nvtop -s | grep "device_name"
   "device_name": "AMD Ryzen 7 7700X 8-Core Processor",

It should also be possible to force it with cardwire launch --gpu=1 but it looks like i forgot to implement it entirely, it will be fixed in the next release

[BreiteSeite]

That's awesome with the env variable. I will try this and consider this closed, as it is exactly what i needed.

Btw. when scrolling through the docs, i briefly saw the environment variable, but i did not realize the 1 is the GPU-ID, i thought it was somehow a boolean value. Maybe it makes sense to rename this to CARDWIRE_FORCE_GPU_ID= or something along the lines?

Thanks so much for the helpful response!

P.S.: How stable is the ID? Would it make more sense use the PCI-ID or something like this?

[BreiteSeite]

For anyone somehow coming across this, if one is using podman quadlets, it's important to add the desired environment variable (Environment=CARDWIRE_FORCE_GPU=1) to the [Service] section (so podman can see/access the GPU as well as the [Container] section so that the process inside the container also has the env variable.

This is great stuff! Really. Thanks so much for creating this awesome tool. In my opinion using the bpf filters is a really good idea for this use case.

[BreiteSeite]

It should also be possible to force it with cardwire launch --gpu=1 but it looks like i forgot to implement it entirely, it will be fixed in the next release

Docs btw. say this:

Launching on a blocked or unavailable GPU is refused with a clear error, switch to Smart mode or unblock the GPU first.

[luytan]

i did forgot to update the docs
And to answer your question, i think i could make it work with a pci id, i will look into it