nbd-client host [ port ] nbd-device [ -connections num ] [ -sdp ] [ -swap ] [ -persist ] [ -nofork ] [ -nonetlink ] [ -systemd-mark ] [ -block-size block size ] [ -timeout seconds ] [ -name name ] [ -certfile certfile ] [ -keyfile keyfile ] [ -cacertfile cacertfile ] [ -tlshostname hostname ]
nbd-client -unix path nbd-device [ -connections num ] [ -sdp ] [ -swap ] [ -persist ] [ -nofork ] [ -nonetlink ] [ -systemd-mark ] [ -block-size block size ] [ -timeout seconds ] [ -name name ]
nbd-client -d nbd-device
nbd-client -c nbd-device
nbd-client -l host [ port ]
nbd-client [ -netlink ] -l host
With nbd-client, you can connect to a server running nbd-server, thus using raw diskspace from that server as a blockdevice on the local client.
To do this, support from the Linux Kernel is necessary, in the form of the Network Block Device (NBD). When you have that, either in the kernel, or as a module, you can connect to an NBD server and use its exported file through a block special file with major mode 43.
Optionally, long options can also be specified with two leading dashes.
The following options are supported:
- -block-size block size
Use a blocksize of "block size". Default is 1024; allowed values are either 512, 1024, 2048 or 4096
- -connections num
Use num connections to the server, to allow speeding up request handling, at the cost of higher resource usage on the server. Use of this option requires kernel support available first with Linux 4.9.
The hostname or IP address of the machine running nbd-server. Since 2.9.15, the NBD utilities support IPv6.
- -timeout seconds
Set the connection timeout to "seconds". For this to work, you need a kernel with support for the NBD_SET_TIMEOUT ioctl; this was introduced into Linus' tree on 2007-10-11, and will be part of kernel 2.6.24.
The TCP port on which nbd-server is running at the server.
The port number defaults to 10809, the IANA-assigned port number for the NBD protocol.
Previous versions of the nbd tools supported an older version of the negotiation protocol known as "oldstyle". This protocol version is no longer supported as of version 3.11 of the nbd support tools.
The block special file (/dev entry) which this nbd-client should connect to, specified as a full path.
When the mode is used wherein no hostname or export name is specified, nbd-client will look up the necessary configuration in the nbdtab file. For more information, see nbdtab(5).
Check whether the specified nbd device is connected.
If the device is connected, nbd-client will exit with an exit state of 0 and print the PID of the nbd-client instance that connected it to stdout.
If the device is not connected or does not exist (for example because the nbd module was not loaded), nbd-client will exit with an exit state of 1 and not print anything on stdout.
If an error occurred, nbd-client will exit with an exit state of 2, and not print anything on stdout either.
Disconnect the specified nbd device from the server
Ask the server for a list of available exports. If the server is exporting over IPv6 as well as over IPv4, this will list all exports twice; otherwise, it should list them all only once.
Note that this option only works with nbd-server processes running version 3.1 or above, and must be enabled in server configuration (with the "allowlist" option) before it can be used.
Starting with version 3.17, nbd-client will default to using the netlink interface to configure an NBD device. This option allows to use the older ioctl() interface to configure the device.
This option is only available if nbd-client was compiled against libnl-genl. If that is not the case, nbd-client will only be able to use the ioctl interface (and the option will not be available).
Note that a future version of nbd-client will require the use of netlink, but it has not yet been decided when that will be the case.
When this option is specified, nbd-client will immediately try to reconnect an nbd device if the connection ever drops unexpectedly due to a lost server or something similar.
Connect to the server using the Socket Direct Protocol (SDP), rather than IP. See nbd-server(5) for details.
Specifies that this NBD device will be used as swapspace. This option attempts to prevent deadlocks by performing mlockall() and adjusting the oom-killer score at an appropriate time. It does not however guarantee that such deadlocks can be avoided.
The systemd init system requires that processes which should not be killed at shutdown time be marked appropriately by replacing the first letter of their argv with an '@' sign.
This option will cause nbd-client to do so.
Note that this only works if nbd-client is run from an initrd; i.e., systemd will ignore such a mark if run from a systemd unit file or from the command line.
Specifies that the NBD client should not detach and daemonize itself. This is mostly useful for debugging.
Note that nbd-client will still fork once to trigger an update to the device node's partition table. It is not possible to disable this.
Disable the use of the NBD_OPT_GO protocol message, and force the use of NBD_OPT_EXPORT_NAME instead.
The NBD protocol has two phases: the negotiation phase, and the transmission phase. To move from negotation to transmission, older clients sent the NBD_OPT_EXPORT_NAME message, for which the server could not produce an error message in case the export name did not exist (or the client had insufficient permissions to access it). Due to those limitations, a replacement message NBD_OPT_GO was created instead, which allows the server to reply with an error in case of any problems.
The protocol allows for a server to discard a message which it does not understand; however, unfortunately some implementations (including older versions of nbd-server) did not handle that situation correctly and would get out of sync with the client when it sent a message which the server did not understand.
When sending NBD_OPT_GO, nbd-client will try to do the right thing and fall back to NBD_OPT_EXPORT_NAME. However, when the server has the above-described bug, then this does not work. In such a situation, the client will issue a diagnostic suggesting the use of this option.
Note that there is a corresponding option for nbdtab, too.
Specifies the name of the export that we want to use. If not specified, nbd-client will ask for a "default" export, if one exists on the server.
Connect to the server over a unix domain socket at path, rather than to a server over a TCP socket. The server must be listening on the given socket.
- -certfile file
Use the specified file as the client certificate for TLS authentication to the server.
- -keyfile file
Use the specified file as the private key for the client cerificate.
- -cacertfile file
Use the specified file as the CA certificate for TLS authentication to the server.
- -tlshostname hostname
Use the specified hostname for the TLS context. If not specified, the hostname used to connect to the server will be used.
Enabling any of the TLS-related options causes the client to use the NBD_OPT_STARTTLS command to upgrade the connection to TLS. Since negotiating TLS support from userspace for a kernel socket would be very involved (if passing keys to kernel space were even possible, which it isn't), the way this is implemented is that the nbd-client process creates a socketpair, one side of which it hands to the kernel, and the other side of which is handed to an encrypting/decrypting proxy. This has the effect that all communication will be encrypted before being sent over the wire; however, doing so is not safe in combination with swapping over an NBD device:
In order to free memory by swapping, the kernel needs to be sure that the write to the nbd device has finalized. For this, it needs to be able to receive an NBD_CMD_WRITE reply which informs it that the write has completed successfully and that the memory may be released. Receiving data over the network, however, requires that the kernel allocate memory first, which is impossible if we're low on memory (a likely situation when trying to swap). This is likely to cause a deadlock when we're low on memory and there are high amounts of network traffic.
To remedy this situation, the kernel sets the PF_MEMALLOC option on the nbd socket; when low on memory, it will throw away all packets except for those destined to a socket with that option set, relying on the normal TCP retransmit system to ensure that data is not lost. This avoids the deadlock described above.
However, the PF_MEMALLOC option is set on the socket that is connected to the nbd device, not the encrypted socket connected to the encrypting/decrypting proxy. As such, when using TLS, the PF_MEMALLOC option is not set on the socket that actually receives data from the network, which means that the deadlock reappears.
For this reason, if the -swap option is used when TLS is in use, nbd-client will issue an appropriate warning.
Some examples of nbd-client usage:
To connect to a server running on port 2000 at host "server.domain.com", using the client's block special file "/dev/nbd0":
nbd-client server.domain.com 2000 /dev/nbd0
To connect to a server running on port 2001 at host "swapserver.domain.com", using the client's block special file "/dev/nbd1", for swap purposes:
nbd-client swapserver.domain.com 2001 /dev/nbd1 -swap
To disconnect the above connection again (after making sure the block special file is not in use anymore):
nbd-client -d /dev/nbd1
The NBD kernel module and the NBD tools have been written by Pavel Macheck (email@example.com).
The kernel module is now maintained by Paul Clements (Paul.Clements@steeleye.com), while the userland tools are maintained by Wouter Verhelst (firstname.lastname@example.org)
This manual page was written by Wouter Verhelst (<email@example.com>) for the Debian GNU/Linux system (but may be used by others). Permission is granted to copy, distribute and/or modify this document under the terms of the GNU General Public License, version 2, as published by the Free Software Foundation.
nbdfuse(1), nbdkit-loop(1), nbdkit-plugin(3), nbd-server(1), nbd-server(5), nbdtab(5).