launching CUDA kernel from a linux kernel?

Is it possible to launch a CUDA kernel from a driver or some other piece of code running inside the linux kernel?

It doesn’t seem like this is possible, as the CUDA runtime library is not linked into the kernel, only the applications. But it seems like being able to do this from a driver for crypto offload or some other processing would be REALLY nice.

So, can this be done? Does it require using some other low level API? If it can be done, how?

To the best of my knowledge, the CUDA API is user space only. There is nothing exposed at the kernel level that would allow another driver access to the hardware.