Build a functionnal librealsense2 with its cuda support

[aurelien-enchanted-tools]

Hello all,

yocto version: langdale
meta-tegra version: langdale
meta-intel-realsense version: langdale
Jetson board: Xavier Nx

To add cuda support of the librealsense2 for the Intel RealSense D435I, I writed a recipes-support/librealsense/librealsense2_2.53.1.bbappend:

PACKAGECONFIG += "\
    cuda \
    rsusb \
    "

inherit cuda
PACKAGECONFIG[cuda] = "-DBUILD_WITH_CUDA:BOOL=ON, -DBUILD_WITH_CUDA:BOOL=OFF, ${CUDA_DEPENDS}"

It builds. But with my application bases on opencv, build with the sdk, which runs with the librealsense, I obtain after several minutes these following errors and an abort:

[…]
NVMAP_IOC_GET_FD failed: Bad address
NVMAP_IOC_GET_FD failed: Bad address
NVMAP_IOC_GET_FD failed: Bad address
NVMAP_IOC_GET_FD failed: Bad address
PosixMemGetSize:122 FD from Handle failed : Bad address
NVMAP_IOC_GET_FD failed: Bad address
[…]

Dmesg logs:

[…]
[ 2247.143376] nvmap: __nvmap_dmabuf_fd() NvMap: FD limit is crossed for uid 1000
[ 2247.176545] nvmap: __nvmap_dmabuf_fd() NvMap: FD limit is crossed for uid 1000
[ 2247.410582] nvmap: __nvmap_dmabuf_fd() NvMap: FD limit is crossed for uid 1000
[ 2247.443816] nvmap: __nvmap_dmabuf_fd() NvMap: FD limit is crossed for uid 1000
[…]

I don’t know where it comes from, why, and how to investigate it. Any ideas ? Maybe something missing ? Thank you in advance.

[kekiefer]

Is your application responsible for allocating buffers? There is a file descriptor leak somewhere, and the NV memory allocator is complaining that there are too many. You can check ulimit to see if the limit is unreasonably low, but likely you'll have to track down the leak.