If a compute is started up while an evacuation of an instance from this
host is still in progress then the destroy_evacuated_instances call will
try to check if the instance is on shared storage to decide if the local
disk needs to deleted from the source node or not. However this call
uses the instance.host to target the RPC call. If the evacuation is
still ongoing then the instance.host might still be set to the source
node. This means the source node during init_host tries to call RPC
on itself. This will always time out as the RPC server is only started
after init_host. Also it is wrong as the shared storage check RPC
should be called on another host. Moreover when this wrongly routed RPC
times out the source compute logs the exception, ignores it, and the
assume the disk is on shared storage so won't clean it up. This means
that a later evacuation of this VM targeting this node will fails as the
instance directory is already present on the node.
The fix is simple, the destroy_evacuated_instances call should always
send the shared storage check RPC call to the destination node of the
evacuation based on the migration record. It will be correct even if the
evacuation is still in progress or even if it is already finished.
Closes-Bug: #2085975
Change-Id: If5ad213649d68da995dad146f0a0c3cacc369309
Balazs Gibizer
2c76fd3baf