Transports
The zrepl RPC layer uses transports to establish a single, bidirectional data stream between an active and passive job.
On the passive (serving) side, the transport also provides the client identity to the upper layers:
this string is used for access control and separation of filesystem sub-trees in sink jobs.
Transports are specified in the connect
or serve
section of a job definition.
Attention
The client identities must be valid ZFS dataset path components
because the sink job uses ${root_fs}/${client_identity}
to determine the client’s subtree.
tcp
Transport
The tcp
transport uses plain TCP, which means that the data is not encrypted on the wire.
Clients are identified by their IPv4 or IPv6 addresses, and the client identity is established through a mapping on the server.
This transport may also be used in conjunction with network-layer encryption and/or VPN tunnels to provide encryption on the wire. To make the IP-based client authentication effective, such solutions should provide authenticated IP addresses. Some options to consider:
WireGuard: Linux-focussed, in-kernel TLS
OpenVPN: Cross-platform VPN, uses tun on *nix
IPSec: Properly standardized, in-kernel network-layer VPN
spiped: think of it as an encrypted pipe between two servers
SSH
sshuttle: VPN-like solution, but using SSH
SSH port forwarding: Systemd user unit & make it start before the zrepl service.
Serve
jobs:
- type: sink
serve:
type: tcp
listen: ":8888"
listen_freebind: true # optional, default false
clients: {
"192.168.122.123" : "mysql01",
"192.168.122.42" : "mx01",
"2001:0db8:85a3::8a2e:0370:7334": "gateway",
# CIDR masks require a '*' in the client identity string
# that is expanded to the client's IP address
"10.23.42.0/24": "cluster-*"
"fde4:8dba:82e1::/64": "san-*"
}
...
listen_freebind
controls whether the socket is allowed to bind to non-local or unconfigured IP addresses (Linux IP_FREEBIND
, FreeBSD IP_BINDANY
).
Enable this option if you want to listen
on a specific IP address that might not yet be configured when the zrepl daemon starts.
Connect
jobs:
- type: push
connect:
type: tcp
address: "10.23.42.23:8888"
dial_timeout: # optional, default 10s
...
tls
Transport
The tls
transport uses TCP + TLS with client authentication using client certificates.
The client identity is the common name (CN) presented in the client certificate.
It is recommended to set up a dedicated CA infrastructure for this transport, e.g. using OpenVPN’s EasyRSA. For a simple 2-machine setup, mutual TLS might also be sufficient. We provide copy-pastable instructions to generate the certificates below.
The implementation uses Go’s TLS library. Since Go binaries are statically linked, you or your distribution need to recompile zrepl when vulnerabilities in that library are disclosed.
All file paths are resolved relative to the zrepl daemon’s working directory. Specify absolute paths if you are unsure what directory that is (or find out from your init system).
If intermediate CAs are used, the full chain must be present in either in the ca
file or the individual cert
files.
Regardless, the client’s certificate must be first in the cert
file, with each following certificate directly certifying the one preceding it (see TLS’s specification).
This is the common default when using a CA management tool.
Note
As of Go 1.15 (zrepl 0.3.0 and newer), the Go TLS / x509 library requrires Subject Alternative Names be present in certificates. You might need to re-generate your certificates using one of the two alternatives provided below.
Note further that zrepl continues to use the CommonName field to assign client identities. Hence, we recommend to keep the Subject Alternative Name and the CommonName in sync.
Serve
jobs:
- type: sink
root_fs: "pool2/backup_laptops"
serve:
type: tls
listen: ":8888"
listen_freebind: true # optional, default false
ca: /etc/zrepl/ca.crt
cert: /etc/zrepl/prod.fullchain
key: /etc/zrepl/prod.key
client_cns:
- "laptop1"
- "homeserver"
The ca
field specified the certificate authority used to validate client certificates.
The client_cns
list specifies a list of accepted client common names (which are also the client identities for this transport).
The listen_freebind
field is explained here.
Connect
jobs:
- type: pull
connect:
type: tls
address: "server1.foo.bar:8888"
ca: /etc/zrepl/ca.crt
cert: /etc/zrepl/backupserver.fullchain
key: /etc/zrepl/backupserver.key
server_cn: "server1"
dial_timeout: # optional, default 10s
The ca
field specifies the CA which signed the server’s certificate (serve.cert
).
The server_cn
specifies the expected common name (CN) of the server’s certificate.
It overrides the hostname specified in address
.
The connection fails if either do not match.
Mutual-TLS between Two Machines
However, for a two-machine setup, self-signed certificates distributed using an out-of-band mechanism will also work just fine:
Suppose you have a push-mode setup, with backups.example.com running the sink job, and prod.example.com running the push job. Run the following OpenSSL commands on each host, substituting HOSTNAME in both filenames and the interactive input prompt by OpenSSL:
(name=HOSTNAME; openssl req -x509 -sha256 -nodes \
-newkey rsa:4096 \
-days 365 \
-keyout $name.key \
-out $name.crt -addext "subjectAltName = DNS:$name" -subj "/CN=$name")
Now copy each machine’s HOSTNAME.crt
to the other machine’s /etc/zrepl/HOSTNAME.crt
, for example using scp.
The serve & connect configuration will thus look like the following:
# on backups.example.com
- type: sink
serve:
type: tls
listen: ":8888"
ca: "/etc/zrepl/prod.example.com.crt"
cert: "/etc/zrepl/backups.example.com.crt"
key: "/etc/zrepl/backups.example.com.key"
client_cns:
- "prod.example.com"
...
# on prod.example.com
- type: push
connect:
type: tls
address:"backups.example.com:8888"
ca: /etc/zrepl/backups.example.com.crt
cert: /etc/zrepl/prod.example.com.crt
key: /etc/zrepl/prod.example.com.key
server_cn: "backups.example.com"
...
ssh+stdinserver
Transport
ssh+stdinserver
uses the ssh
command and some features of the server-side SSH authorized_keys
file.
It is less efficient than other transports because the data passes through two more pipes.
However, it is fairly convenient to set up and allows the zrepl daemon to not be directly exposed to the internet, because all traffic passes through the system’s SSH server.
The concept is inspired by git shell and Borg Backup.
The implementation is provided by the Go package github.com/problame/go-netssh
.
Note
ssh+stdinserver
generally provides inferior error detection and handling compared to the tcp
and tls
transports.
When encountering such problems, consider using tcp
or tls
transports, or help improve package go-netssh.
Serve
jobs:
- type: source
serve:
type: stdinserver
client_identities:
- "client1"
- "client2"
...
First of all, note that type=stdinserver
in this case:
Currently, only connect.type=ssh+stdinserver
can connect to a serve.type=stdinserver
, but we want to keep that option open for future extensions.
The serving job opens a UNIX socket named after client_identity
in the runtime directory.
In our example above, that is /var/run/zrepl/stdinserver/client1
and /var/run/zrepl/stdinserver/client2
.
On the same machine, the zrepl stdinserver $client_identity
command connects to /var/run/zrepl/stdinserver/$client_identity
.
It then passes its stdin and stdout file descriptors to the zrepl daemon via cmsg(3).
zrepl daemon in turn combines them into an object implementing net.Conn
:
a Write()
turns into a write to stdout, a Read()
turns into a read from stdin.
Interactive use of the stdinserver
subcommand does not make much sense.
However, we can force its execution when a user with a particular SSH pubkey connects via SSH.
This can be achieved with an entry in the authorized_keys
file of the serving zrepl daemon.
# for OpenSSH >= 7.2
command="zrepl stdinserver CLIENT_IDENTITY",restrict CLIENT_SSH_KEY
# for older OpenSSH versions
command="zrepl stdinserver CLIENT_IDENTITY",no-port-forwarding,no-X11-forwarding,no-pty,no-agent-forwarding,no-user-rc CLIENT_SSH_KEY
CLIENT_IDENTITY is substituted with an entry from
client_identities
in our exampleCLIENT_SSH_KEY is substituted with the public part of the SSH keypair specified in the
connect.identity_file
directive on the connecting host.
Note
You may need to adjust the PermitRootLogin
option in /etc/ssh/sshd_config
to forced-commands-only
or higher for this to work.
Refer to sshd_config(5) for details.
To recap, this is of how client authentication works with the ssh+stdinserver
transport:
Connections to the
/var/run/zrepl/stdinserver/${client_identity}
UNIX socket are blindly trusted by zrepl daemon. The connection client identity is the name of the socket, i.e.${client_identity}
.Thus, the runtime directory must be private to the zrepl user (this is checked by zrepl daemon)
The admin of the host with the serving zrepl daemon controls the
authorized_keys
file.Thus, the administrator controls the mapping
PUBKEY -> CLIENT_IDENTITY
.
Connect
jobs:
- type: pull
connect:
type: ssh+stdinserver
host: prod.example.com
user: root
port: 22
identity_file: /etc/zrepl/ssh/identity
# options: # optional, default [], `-o` arguments passed to ssh
# - "Compression=yes"
# dial_timeout: 10s # optional, default 10s, max time.Duration until initial handshake is completed
The connecting zrepl daemon
Creates a pipe
Forks
In the forked process
Replaces forked stdin and stdout with the corresponding pipe ends
Executes the
ssh
binary found in$PATH
.The identity file (
-i
) is set to$identity_file
.The remote user, host and port correspond to those configured.
Further options can be specified using the
options
field, which appends each entry in the list to the command line using-o $entry
.
Wraps the pipe ends in a
net.Conn
and returns it to the RPC layer.
As discussed in the section above, the connecting zrepl daemon expects that zrepl stdinserver $client_identity
is executed automatically via an authorized_keys
file entry.
The known_hosts
file used by the ssh command must contain an entry for connect.host
prior to starting zrepl.
Thus, run the following on the pulling host’s command line (substituting connect.host
):
ssh -i /etc/zrepl/ssh/identity root@prod.example.com
Note
The environment variables of the underlying SSH process are cleared. $SSH_AUTH_SOCK
will not be available.
It is suggested to create a separate, unencrypted SSH key solely for that purpose.
local
Transport
The local transport can be used to implement local replication, i.e., push replication between a push and sink job defined in the same configuration file.
The listener_name
is analogous to a hostname and must match between serve
and connect
.
The client_identity
is used by the sink as documented above.
jobs:
- type: sink
serve:
type: local
listener_name: localsink
...
- type: push
connect:
type: local
listener_name: localsink
client_identity: local_backup
dial_timeout: 2s # optional, 0 for no timeout
...