Capture card vendor try to implement V4L2 driver with Nvidia RDMA

[david242]

Hello,

I am a developer of captured card vendor.

We have a problem when enabling GPUDirect RDMA on Ubuntu.

Environment:
OS: Ubuntu 20.04 x86_64
kernel version: 5.15.0-52-generic
Nvidia driver: 520.56.06

Kernel rejects to load of our driver because the driver uses GPL symbols and proprietary symbols.

sudo modprobe LXV4L2D_SC0710
dmesg
[ 9217.151262] LXV4L2D_SC0710: module using GPL-only symbols uses symbols from proprietary module nvidia.
[ 9217.151267] LXV4L2D_SC0710: Unknown symbol nvidia_p2p_dma_unmap_pages (err -2)
[ 9217.151330] LXV4L2D_SC0710: module using GPL-only symbols uses symbols from proprietary module nvidia.
[ 9217.151332] LXV4L2D_SC0710: Unknown symbol nvidia_p2p_get_pages (err -2)
[ 9217.151359] LXV4L2D_SC0710: module using GPL-only symbols uses symbols from proprietary module nvidia.
[ 9217.151361] LXV4L2D_SC0710: Unknown symbol nvidia_p2p_put_pages (err -2)
[ 9217.151381] LXV4L2D_SC0710: module using GPL-only symbols uses symbols from proprietary module nvidia.
[ 9217.151382] LXV4L2D_SC0710: Unknown symbol nvidia_p2p_dma_map_pages (err -2)
[ 9217.151399] LXV4L2D_SC0710: module using GPL-only symbols uses symbols from proprietary module nvidia.
[ 9217.151400] LXV4L2D_SC0710: Unknown symbol nvidia_p2p_free_dma_mapping (err -2)
[ 9217.151404] LXV4L2D_SC0710: module using GPL-only symbols uses symbols from proprietary module nvidia.
[ 9217.151405] LXV4L2D_SC0710: Unknown symbol nvidia_p2p_free_page_table (err -2)

Captured card using V4L2 framework which is GPL and nvidia_p2p_* are proprietary.

What are the best practices to solve this problem? 

If using open source Nvidia driver, some old Quadro cards don't on the supportting list.

BR,
David

[ylazimy]

Hello David,
In order to help you, we will need more information on the capture card manufacturer and model. Can you please provide more context?
thanks!

[jinl]

Hi David, You could update your stack to R515+ through GitHub - NVIDIA/open-gpu-kernel-modules: NVIDIA Linux open GPU kernel module source. Holoscan OpenEmbedded recipes release v0.3.0 solves the issue with enabling RDMA by using OpenRM r515 drivers, please take a look at GitHub - NVIDIA/meta-tegra-clara-holoscan-mgx at v0.3.0.
Regarding If using open source Nvidia driver, some old Quadro cards don't on the supportting list. : could you tell us a little more about which hardware you’re using?

[david242]

Hi Ylazimy & Jinl,

Thanks for your quick reply.

Let me re-introduce myself, I am the engineer of YUAN High-Tech and we are the partner of Nvidia GPUDirect.
Because the FAE of Nvidia in Taiwan can't find a suitable person for this question, he suggested I post on this forum directly.

Many customers want to use Nvidia GPUDirect to improve performance and latency.
And thank GPUDirect and Holoscan SDK, we got 1-2 frames latency (4k 60fps, about 16-32ms) on Clara AGX Developer Kit.

I successfully used R515 open-source driver on a PC and enabled Nvidia GPUDirect.
Next, I will try to use R515 open-source driver on Clara AGX Developer Kit.
Because I’m not familiar with OpenEmbedded/Yocto, this will take more time to finish.

For now, has two customers
Customer A
Platform: Clara AGX Developer Kit.
OS: Holopack 1.1, l4t 34.1.2
GPU: Quadro 6000
Using Holoscan SDK: Yes
Captured card: SC0710 HDMI or SDVoE (https://www.yuan.com.tw/products/capture/4k/sc710n1_l_hdmi2.htm)

Customer B
Platform: x86 PC
OS: Ubuntu 20.04
GPU: Quadro A6000
Using Holoscan SDK: Unknow, they don’t ask.
Captured card: SC0710 HDMI

Could you tell us a little more about which hardware you’re using
– This is just because I have a very old Quadro which is M4000, I think most customers will use the newer Quadro.
So just forget this.

Thank you for your help. If I meet a new problem, I will post a new question.

BR,
david