Wrongly using /buckets/<volumeId > when path is specified in volume

[Ruakij]

Problem

By default, the driver will use the volumeId to create a new folder under /buckets/<volumeId>. This also gets deleted when the volume is deleted.

But as we can override the path given to the filer by setting spec.csi.volumeAttributes.path, this doesnt work then.

Behavior

• Create PV
  • Folder is not created
• Delete PVC attached to PV (with policy Delete)
  • Error because folder doesnt exist. PV stays marked with "Error" and has to be removed manually
• Mounting
  • Mounter doesnt complain the filer.path doesnt exist. Files can be created, etc, but are inaccessible from parent-directories. (Bug in weed mount or filer to allow this?)

Expected Behaviour

• When creating a volume:
  • specified path is created if it didnt exist
• When deleting a volume:
  • specified path is deleted
    • I'd propose having a flag to not create/delete anything as this can be dangerous and unwanted when provisioned statically
• Mounting
  • When filer.path doesnt exist, error. (or auto-create?)
    • This can happen when the folder was deleted after the Volume is created.

Steps to recreate

1. Create PersistentVolume with custom path

apiVersion: v1
kind: PersistentVolume
metadata:
  name: test
spec:
  capacity:
    storage: 20Gi
  csi:
    driver: seaweedfs-csi-driver
    volumeHandle: test
    volumeAttributes:
      path: /mytest
  accessModes:
    - ReadWriteOnce
  persistentVolumeReclaimPolicy: Delete
  storageClassName: seaweedfs
  volumeMode: Filesystem

2. Create PersistentVolumeClaim to bind PersistentVolume

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: test
spec:
  resources:
    requests:
      storage: 20G
  volumeMode: Filesystem
  accessModes:
    - ReadWriteOnce
  storageClassName: seaweedfs

3. Check filer, see no path /mytest is created
   Interestingly, also the path /buckets/test doesnt exist.
4. Delete PVC
5. See PV being marked as "Error" in Kubernetes. (not sure if the driver-log also contains errors)

---

CreateVolume:

seaweedfs-csi-driver/pkg/driver/controllerserver.go

Line 53 in d777b87

if err := filer_pb.Mkdir(cs.Driver, "/buckets", volumeId, nil); err != nil {

DeleteVolume:

seaweedfs-csi-driver/pkg/driver/controllerserver.go

Line 83 in d777b87

if err := filer_pb.Remove(cs.Driver, "/buckets", volumeId, true, true, true, false, nil); err != nil {

ValidateVolumeCapabilities:

seaweedfs-csi-driver/pkg/driver/controllerserver.go

Line 127 in d777b87

exists, err := filer_pb.Exists(cs.Driver, "/buckets", req.GetVolumeId(), true)

---

Depending on what we want to allow, we could ideally get the path of the volume and operate on that. This would then be equivalent to volumes which dont have a path set.
Though i am not sure how hard it is to get that information as its not passed with the Request.

From what i observed, the CreateVolume isnt even called when a PV is created, only when it happens through a PVC?
Maybe this is an expected behaviour?

[n9]

What I have found yet:

• path was introduced in 5c2298b, it replaced older bucketName
• but it seems to me that the same issue was with the bucketName

I seems to me that path should be also handled in controllerserver, not just in nodeserver (via mounter).

@kvaster, are you using path? In what case it is working for you?

[kvaster]

There are at least two ways of using seaweedfs:

• Create path (PV) automatically - RWO. Here we need to create / allocate bucket and to delete it later.
• Use common/shared path - RWX. Here we need just a path and we really don't need to delete this path.

I will recheck logic and will get back.