On Wed, 03/16 11:27, Paolo Bonzini wrote: > > > On 16/03/2016 11:10, Fam Zheng wrote: > > An empty begin/end pair is almost the same as a bare bdrv_drain except > > the aio_poll inside is wrapped by > > aio_disable_external/aio_enable_external. > > > > This is safer, and is the only way to achieve quiescence in this > > aio_poll(), because bdrv_drained_begin/end pair cannot span across > > context detach/attach options, so it's not possible to do by the caller. > > I'm still not sure about this patch. > > When starting dataplane, the ioeventfd is registered with iohandler.c so > bdrv_drained_begin/end is not necessary.
You are right, and looks like the k->set_host_notifier() above blk_set_aio_context would disable the fd anyway. > > Likewise when stopping dataplane bdrv_set_aio_context is called after > the thread has been stopped and thus the ioeventfd is not registered > anymore as an external client. Right. Fam
