cudaHostRegister alternative for already pinned memory

Our input buffers are pinned and should remain so at all cost. When we use cudaHostRegister to register them, the cuda driver tries to pin them again, and the corresponding unregister will make them pageable again, which is not desired.

We are looking for a way to register a memory range with the cuda driver (so it knows it can use DMA), without the driver controlling the pinning. Any ideas ?