On 10/13/2016 07:34 PM, Paolo Bonzini wrote:
> This patch reorganizes aio_poll callers to establish new rules for
> dataplane locking. The idea is that I/O operations on a dataplane
> BDS (i.e. one where the AioContext is not the main one) do not call
> aio_poll anymore. Instead, they wait for the operation to end in the
> other I/O thread, at which point the other I/O thread calls bdrv_wakeup
> to wake up the main thread.
> With this change, only one thread runs aio_poll for an AioContext.
> While aio_context_acquire/release is still needed to protect the BDSes,
> it need not interrupt the other thread's event loop anymore, and therefore
> it does not need contention callbacks anymore. Thus the patch can remove
> RFifoLock. This fixes possible hangs in bdrv_drain_all, reproducible (for
> example) by unplugging a virtio-scsi-dataplane device while there is I/O
> going on for a virtio-blk-dataplane on the same I/O thread.
Have you seen improvements or deteriorations in performance with single disks
or multiple disks? Is there a branch?