What problem are we solving?

When we use the s3fs mounter, even if multiple pods use the same volume, the driver will start a process for each of them to mount the volume. This is wasteful of resources and bad for management.

o ps -ef | grep s3fs | grep pod
root     26654  3594  0 15:47 ?        00:00:00 s3fs custom / /var/lib/kubelet/pods/7e72ff6c-8096-4de3-973f-9d1a97fc3a89/volumes/kubernetes.io~csi/pvc-1a2313d1-ae74-45ee-bfa4-63106d7084ec/mount -o use_path_request_style -o url=http://10.134.80.223:9000 -o endpoint= -o allow_other -o mp_umask=000
root     34834  3594  0 19:32 ?        00:00:00 s3fs custom / /var/lib/kubelet/pods/5b8a4cb7-3897-42eb-94cf-049d0485a11b/volumes/kubernetes.io~csi/pvc-1a2313d1-ae74-45ee-bfa4-63106d7084ec/mount -o use_path_request_style -o url=http://10.134.80.223:9000 -o endpoint= -o allow_other -o mp_umask=000

How are we solving the problem?

The idea is that pods using the same volume need to share a mount process.
We currently solve this problem based on the CSI specification.

1. When a volume is first used on a node, NodeStageVolume will start a process to mount it to host's /var/lib/kubelet/plugins/kubernetes.io/csi/pv/<pvc>/globalmount directory.
2. Later when other pod uses the volume, NodePublishVolume just needs to bind mount /var/lib/kubelet/plugins/kubernetes.io/csi/pv/<pvc>/globalmount to /var/lib/kubelet/pods/<pod>/volumes/kubernetes.io~csi/<pvc>/mount directory.
3. If the pod using the volume on the node is deleted, then NodeUnpublishVolume only needs to unmount /var/lib/kubelet/pods/<pod>/volumes/kubernetes.io~csi/<pvc>/mount.
4. Finally, when the volume is not used by any pods on the node, NodeUnstageVolume will unmount the volume from /var/lib/kubelet/plugins/kubernetes.io/csi/pv/<pvc>/globalmount directory.

We made a significant change to the code to make it easier to perform similar optimizations for other mounters.

Additionally, according to the specification, we also consider scenarios for handling concurrent calls.
https://github.com/container-storage-interface/spec/blob/master/spec.md#concurrency