RFC qemu-coroutine: use a ring per thread for the pool

This series removes the need for locking the coroutine pool. During collection of perf data it showed up that there is serious time spend in locking the coroutine pool during heavy I/O from the vServer.

This also needs serious benchmarking. Using Kevins qemu-img bench there is a 3-4% speedup. Inside vServers it seems to heavily dependend on the individual CPU and SMP configuration.

I just wanted to post these patches here to discuss if this is an idea that we should look futher into.

KNOWN ISSUES:
- the I/O thread qobject is not destroyed on qemu termination. These seems to be the case for all qobjects?! Therefore the per thread coroutine pool is not cleaned up.
- without ioeventfd there is effectively no longer a coroutine pool since the coroutines are created in a different thread then they are freed.

Peter Lieven (3):
Revert "coroutine: make pool size dynamic" block/block-backend.c: remove coroutine pool reservation
qemu-coroutine: use a ring per thread for the pool

 block/block-backend.c     |    6 ----
 include/block/coroutine.h |   13 +-------
 iothread.c                |    3 ++
 qemu-coroutine.c          |   78 +++++++++++++++------------------------------
 4 files changed, 30 insertions(+), 70 deletions(-)

1.7.9.5