OpenStack Mitaka brought the support of a new feature. This feature is a follow-up of the Nova discard implementation. Now it is possible to configure Cinder per backend.

It is mandatory to use a Glance image with these two properties hw_scsi_model=virtio-scsi and hw_disk_bus=scsi.

If your image does not have the right property just update it with this command:

$ glance image-update <image id > --property hw_scsi_model=virtio-scsi --property hw_disk_bus=scsi

As an operator, in order for your backend to report discard capabilities, you need to configure Cinder volume. It is pretty simple so simply append the following line report_discard_supported = True in your backend section:

[ceph]
rbd_max_clone_depth = 5
rbd_flatten_volume_from_snapshot = False
rbd_uuid = 6ed23bd0-aa89-4b0b-b5c6-77bf4da08afd
rbd_user = cinder
rbd_pool = volumes
rbd_ceph_conf =
volume_driver = cinder.volume.drivers.rbd.RBDDriver
volume_backend_name = ceph
report_discard_supported = true

Eventually restart your Cinder volume service. Now attach a block device. If you want to validate that the device reports discard, you can check the libvirt xml file of the virtual machine. Alternatively within the virtual machine you can check /sys/block/sdb/queue/discard_*.

Always be careful when using fstrim after forward to reclaim space has this will introduce some performance penalty.