How GPU memory got allocated?

[lrh12580]

Hi there, I had a question about how GPU memory got allocated.

cudaMalloc() is the API interface for CUDA programs to reserve GPU memory in user programs. In the kernel space, based on my understanding, the user space memory allocation is processed in two phases.

(1) ioctl() syscall to reserve GPU memory.
(2) mmap() syscall to map the reserved memory to user's virtual memory space.

Wondering if my understanding is correct. If so, could anyone point me to which ioctl() handler should I take a look at?

Appreciate it for the help.