"Daniel P. Berrange" <berra...@redhat.com> writes: > The only other alternative would be for the destination to open the disks, > but not immediately acquire the locks. In the final stage of migration have > the source release its lock & signal to the dest that it can now acquire > the lock. The assumption being that the lock acquisition is far less likely > to fail than the open(), so we focus on making sure we can properly handle > open() failure.
This is what we've implemented as a prototype for sheepdog, with bdrv_claim (happens on vm_start) and bdrv_release (happens on exit/vm_stop) hooks available to block drivers. Cheers, Chris.
