GPUDirect RDMA - Multithreading

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
,
1

Hello,

I am writing a driver which uses the GPUDirect RDMA on the Jetson for direct access to CUDA-Memory. I have a number of distinct data streams, which are transfered via DMA. The different streams are managed by different threads (in the userland). Via fileopts I can manage these transfers.
Can I call the interface (see below) from different threads without additional precautions, like mutex or spinlocks? Or do I need some sort of synchronization? Due to the architecture the calls for one data-region/stream are always sequential. But I could e.g. have two parallel calls to nvidia_p2p_dma_map_pages for different memory regions.

I am using the following functions:

  • nvidia_p2p_free_page_table
  • nvidia_p2p_put_pages
  • nvidia_p2p_get_pages
  • nvidia_p2p_dma_map_pages
  • nvidia_p2p_dma_unmap_pages

Best regards and thanks in advance,
Gerrit

3

Our team is investigating now, will do the update soon.

Thanks

4

By checking with internal team ,they said it should be ok to call these parallely for different memory regions.
For additional info. look at the sample gdrcopy/gdrdrv.c at master · NVIDIA/gdrcopy · GitHub.

5

Thank you very much for the clarification and link to the corresponding source file.