On an RTX 5090 using the NVIDIA open kernel module, exporting GPU memory as a DMA-BUF fails:
- Driver: 595.84
- CUDA: 13.2
- Kernel: 7.0.0-30-generic
- Memory allocated directly with the CUDA Driver API using cuMemAlloc()
- CU_DEVICE_ATTRIBUTE_DMA_BUF_SUPPORTED returns 0
- cuMemGetHandleForAddressRange(…, CU_MEM_RANGE_HANDLE_TYPE_DMA_BUF_FD, 0) returns CUDA_ERROR_NOT_SUPPORTED (801)
Is CUDA DMA-BUF export intentionally unsupported on the GeForce RTX 5090 or RTX 50 series? If so, which GPU families support it? If it should work, is any configuration required beyond using the open kernel module?