Calling VirtualQuery on a buffer allocated by cudaMallocHost reports that the buffer’s type is MEM_PRIVATE. Is there a reason why the buffer type isn’t set to MEM_MAPPED (as it should, I guess)?
Maybe only the guys from the CUDA team can answer this one… Q: is it OK to use a buffer allocated by cudaMallocHost in the context of a device driver (e.g. doing DMA directly into that buffer)?