On 05/27/2015 01:19 AM, Fam Zheng wrote: > Currently the host notifier is checked by all aio_poll, which is not > safe. For example, in a qmp transaction that takes snapshots or starts > drive-backup on multiple dataplane disks, the atomicity could be broken: > There could be one or more bdrv_drain_all() calls in each transaction > opeartion, which will unconditinally calls one or more aio_poll on the
s/opeartion/operation/ s/unconditinally calls/unconditionally call/ > AioContext: > > qmp transaction > backup A prepare > ... > bdrv_drain_all() > aio_poll(A) > aio_poll(B) > aio_poll(C) > ... > ... > backup B prepare > ... > bdrv_drain_all() > aio_poll(A) > aio_poll(B) > -> aio_poll(C) > ... > snapshot C prepare > ... > bdrv_drain_all() > aio_poll(A) > aio_poll(B) > aio_poll(C) > ... > > If the aio_poll(C) in the middle receives a new virtio-blk request from > ioeventfd, a new request will be submitted to C, then snapshot C is > inconsistent. > > To avoid that, use aio_set_io_event_notifier so the behavior is the same > as in main loop. > > Signed-off-by: Fam Zheng <f...@redhat.com> > --- > hw/block/dataplane/virtio-blk.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > again, letting others comment on the technical merits -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
signature.asc
Description: OpenPGP digital signature
