So, to improve performance, I'd suggest a way that the hardware
   driver can request the next block even when it's not yet done with
   the one it is currently using.

That's certainly a better thing to do, and I'd encourage you to go for
it.  B-)

It's also been suggested to me (by one of the authors of AudioFile)
that the buffer (either fictitious, or, if possible, on the board
itself) should be mmap(2)ed into the process, and system calls should
only be used to request that blocks of data be input or output.  This
would be somewhat more efficient than copyin()/copyout().