$ benji restore --help
usage: benji restore [-h] [-s] [-f] [-d] [-S STORAGE] version_uid destination

positional arguments:
  version_uid           Version UID to restore
  destination           Destination URL

optional arguments:
  -h, --help            show this help message and exit
  -s, --sparse          Restore only existing blocks
  -f, --force           Overwrite an existing file, device or image
  -d, --database-less   Restore without requiring the database
  -S STORAGE, --storage STORAGE
                        Source storage (if unspecified the default is used)

There are two possible restore options with Benji.


If you determined the version you want to restore it is a good idea to protect this version with benji protect to prevent any accidental or automatic removal by any retention policy enforcement that you might have configured.


A restore can write a version to any one of the supported and configured I/O instances. This doesn’t have to be the same I/O instance that was used for creating the backup. The URL syntax is the same as the one used for the backup command. Examples include:

  • Command line: benji restore --sparse V00000001 example-1:///tmp/vm1-image.qcow2

    I/O module configuration:

      - name: example-1
        module: file

    Restores version V00000001 to path /tmp/vm1-image.qcow2 in the local filesystem

  • Command line: benji restore --sparse V00000002 example-2:pool-1/image-2

    I/O module configuration:

      - name: example-2
        module: rbdaio

    Restores version V00000002 to an RBD image named image-2 in the Ceph pool pool-1. The RBD image is created in the process with the configured RBD image features. If no RBD image features are specified in the configuration Ceph’s default is used.


Normally the name chosen for an instance of the file module is file and the name used for an instance of an rbd or rbdaio module is rbd. But that is not required and there can be multiple instances of the same I/O module with different configurations under different names.

If the target already exists, i.e.

  • it is a device file

  • an existing Ceph RBD volume

  • or an existing image file

the --force option needs to be added to the command line. Otherwise Benji won’t overwrite the existing resource.

Generally the usage of the --sparse option is advisable to skip the restore of sparse (empty) blocks. This increases restore performance and also decreases space usage in most cases. When --sparse is not specified sparse blocks are written as blocks of zeros.


If you use --sparse to restore to an existing device or file, sparse blocks will not be written, so whatever random data was in the location of the sparse block before the restore will remain. This is not the case with Ceph RBD as --sparse will discard all currently used blocks before beginning the restore.

Restoring without a database

benji restore also supports a mode to restore a version even when the database is not available. This mode is activated by passing the --database-less switch to benji restore. Benji will import the metadata backup of the specified version from the storage into an ad-hoc in-memory database and then restore the image normally. If you have configured more than one storage location and the version to restore does not reside on the default storage you need to specify the storage location with --storage.

This mode is for failure scenarios where the database is unavailable. But because of this unavailability it is impossible to execute commands like benji ls to determine the right version for the restore. Reports (based on benji ls) need to be generated and saved beforehand so that this information is still available in another way.

NBD Server

$ benji nbd --help
usage: benji nbd [-h] [-a BIND_ADDRESS] [-p BIND_PORT] [-r]

optional arguments:
  -h, --help            show this help message and exit
  -a BIND_ADDRESS, --bind-address BIND_ADDRESS
                        Bind to the specified IP address (default:
  -p BIND_PORT, --bind-port BIND_PORT
                        Bind to the specified port (default: 10809)
  -r, --read-only       NBD device is read-only (default: False)

Benji comes with its own NBD server which when started exports all known versions. The versions can then be mounted on any Linux host. The requirements on the Linux host are:

  • loaded nbd kernel module (modprobe nbd as root)

  • installed nbd-client program (RPM package nbd on RHEL, CentOS and Fedora)

The nbd-client contacts Benji’s NBD server and connects an exported version to am NBD block device (/dev/nbd*) on the Linux host. If the image contains a filesystem it can be mounted normally. You can then search for the relevant files and restore them.


The NBD server can take some time when it accesses a large version (in the order of multiple Terabytes) for the first time. This holds especially true if the version is exported read-write. As a consequence a timeout specified on the nbd-client command line might need to be increased otherwise the connection will time out.

Known Issues with nbd-client

There are some known issues with nbd-client:

  • There have been numerous problems with different versions of nbd-client when used together with Benji. Recent findings (11/18/2019) suggest that a bug in nbd-client introduced in this commit and fixed in this commit was the reason for most of them. Benji versions 0.8.0 and newer contain a workaround for this bug.

  • Some versions of nbd-client (like 3.17) use a timeout value of zero which also leads to problems. Please explicitly specify a timeout with -t in these cases. To make the confusion complete some distributions have back-ported the fix for this issue. So 3.17 on Fedora is actually fine without -t.

  • If the image has a partition table the Linux kernel will have problems parsing the partition table when the device block size used by nbd-client is different from the one used during creation of the partition table. Please use the -b option of nbd-client to specify the original device block size which normally will be 512. Newer versions of nbd-client already changed the default block size from 1024 to 512 because of this. See

Read-Only Mount

This command will run the NBD server in read-only mode and wait for incoming connections:

$ benji nbd -r
    INFO: Starting to serve nbd on

Benji’s NBD server will serve all available versions and it is possible to access each one of them as a block device by using nbd-client und the in-kernel nbd driver:

# Load the nbd kernel module
$ sudo modprobe nbd

# Connect a Benji version to a free NBD block device
$ sudo nbd-client -p -b 512 -t 10 -N V0000000001 10809 /dev/nbd0
Negotiation: ..size = 10MB
bs=512, sz=10485760 bytes

# Detect partitions
# (partprobe might throw a few WARNINGs because we're read-only)
partprobe /dev/nbd0

# Mount the filesystem
mount -o ro /dev/nbd0p1 /mnt

If the image just contains a filesystem without a partition table the partprobe command can be skipped and the /dev/nbd0 device can be mounted directly. Some filesystem will require additional mount options as they try to write to the device even when ro is specified.

The NBD server will signal an incoming connection:

INFO: Incoming connection from
INFO: [] Negotiated export: V0000000001
INFO: nbd is read only.

After you’re done using the filesystem you need to unmount it and disconnect the NBD device:

umount /mnt
nbd-client -d /dev/nbd0

You can then either reconnect to this or another version or terminate Benji’s NBD server. The server is also able to serve multiple versions at once.

Benji’s NBD server by default listens on (i.e. localhost) for incoming connections. For the server to be reachable from the outside bind it to or the specific address of another interface:

benji nbd -a -r


Exposing the NBD server on an external interface has security implications.

Read-Write Mount

In addition to providing read-only access, Benji also allows read-write access in a safe way. This means, the original version will not be modified. To access the available versions in read-write mode start the NBD server without the -r option.

After connecting the NBD device you can initiate any repair procedures required like fsck. Any writes to the device will initiate a copy-on-write (COW) of the original blocks to a new version which is dynamically created by Benji.

After disconnecting the NBD device Benji will start to fixate the COW version. Depending on how many changes have been done to the original version this will take some time!:

INFO: [] disconnecting
INFO: Fixating version V0000000002 with 1024 blocks, please wait!
INFO: Fixation done. Deleting temporary data, please wait!
INFO: Finished.


If you end the NBD server before the last “INFO: Finished.” is reported, your copy-on-write clone will not be written completely and thus be incomplete. However, the original backup version will be untouched in any case.

The newly created version can be seen in the output of benji ls:

$ benji ls
    INFO: $ benji ls
|         date        |     uid     | name | snapshot                                |     size | block_size | valid | protected | tags |
| 2018-06-10T01:00:43 | V0000000001 | test |                                         | 41943040 |    4194304 |  True |   False   |      |
| 2018-06-10T01:01:16 | V0000000002 | test | nbd-cow-V0000000001-2018-06-10T01:01:16 | 41943040 |    4194304 |  True |    True   |      |

The name will be the same as the original version. The snapshot will start with the prefix nbd-cow- followed by the version UID followed by a timestamp.

The COW version will automatically be marked as protected by Benji to prevent removal by any automatic retention policy enforcement configured. This ensures the new version won’t be destroyed accidentally. To be able to remove the version the protection needs to be lifted with benji unprotect.


The new created COW version can be restored just like any other version. Both the original and the COW version are independent from each other and each can be removed without affecting the other.

Restoring Invalid Versions

During a restore Benji will compare each restored block’s checksum to the one stored in the database. This even happens when the block has been marked as invalid previously. If it encounters a difference, the block and all versions also referencing this block will be marked as invalid and a warning will be given:

ERROR: Checksum mismatch during restore for block 9 (UID 1-a) (is: e36cee7fd34ae637... should-be: dea186672147e1e3..., block.valid: True). Block restored is invalid.
 INFO: Marked block invalid (UID 1-a, Checksum dea186672147e1e3. Affected versions: V0000000001, V0000000002
 INFO: Marked version invalid (UID V0000000001)
 INFO: Marked version invalid (UID V0000000002)

Even when encountering such an error, Benji will continue the restore.


The philosophy behind this is that restores should always succeed, even if there is data corruption. Often invalid data is in irrelevant places or can be fixed later. You get as much of your data back as possible!