engines/libaio: fix issue with EAGAIN
If we get -EAGAIN on IO submission, we should back off and reap
some events. Currently we just loop trying to resubmit, and this
has been reported to cause hangs in some virtualized environments.
There's no record of this ever being hit on raw metal.
So implement this backoff. Unfortunately it requires rewriting
the io_u tracking in libaio, since otherwise we depend on being
able to submit all the io_u's we are handed. So change the
tracking to be a ring buffer, with submission bumping the head
and completion bumping the tail. With this in place, we can't
stop issuing IO at anytime, since we'll just continue where
we left off.
Signed-off-by: Jens Axboe <axboe@fb.com>