On start, external Nvidia monitors sometimes don't work (gaze15 GTX 1650)

[ids1024]

This isn't a new problem, and restarting the compositor can fix it. But I hadn't previously looked at nvidia-smi when this is happening:

|    0   N/A  N/A   3922993    C+G   ./target/release/cosmic-comp               3863MiB

There are dmesg errors of:

[drm:nv_drm_gem_alloc_nvkms_memory_ioctl [nvidia_drm]] *ERROR* [nvidia-drm] [GPU ID 0x00000100] Failed to allocate NVKMS memory for GEM object

So somehow cosmic-comp quickly fills vram with allocations until the GPU runs out of memory? Not sure what would cause that...

[Drakulix]

Oh this is fascinating... in a bad way. I suspect we get an error from kms somewhere and cycle in retries or something similar.

Quite hard to debug, it would be incredibly helpful to get figure out, where those allocations happen, but I suspect getting a valgrind run of this will be tricky. I have seen this happening, but it's not exactly frequent.

[ids1024]

It's VRAM that's being leaked here, so I'm not sure Valgrind would help. Apparently CUDA has a memcheck tool, but I don't know if there are any tools that could be used outside of CUDA.

I guess tracing EGL/GBM calls and seeing if something is being called repeatedly and returning an error may be informative.