The dangling volume mechanism we currently use for recovering volumes that are incorrectly attached will not work for CSI.

This mechanism is in use by Cinder and EBS drivers at present - https://github.com/kubernetes/kubernetes/blob/master/pkg/volume/util/operationexecutor/operation_generator.go#L338 and solves real issues where a volume may be erroneously attached to the node.

In past building exact state of the world after controller-manager recovers from a crash was kinda not very precise which results in bugs like this. I am wondering if for CSI we should be able to use VolumeAttachment objects to rebuild the state and workaround/fix problems like this.

/sig storage

cc @jsafrane @msau42 @jingxu97 @davidz627