On Wed, Apr 25, 2007 at 04:24:32AM +0300, Mindaugas R. wrote:
> Hello,
> 
> 
> Shortly about the mechanism: each process may have an AIO-worker thread,
> which is initialized only after the first AIO operation call, and destroyed
> when process exits. All I/O operations are processed in this thread, queue of
> the operations is also per-process. There are no queue of the completed
> operations - state and values are stored in the internal kernel variables of
> aiocb structure (this is allowed by POSIX) for the performance purposes.

Just out of curiosity, if the implementation uses an additional thread in
the user process, why is any kernel support required at all?

Thor