rsync

README

gokrazy rsync

This repository contains a native Go rsync implementation: the gokr-rsync command implements an rsync client and server, which can send or receive files (all directions supported). Daemon mode is supported, meaning you can deploy gokr-rsync behind SSH (anonymous or authorized), as command or daemon, or listening directly on the network (on port 873/tcp by default).

The following known improvements are not yet implemented:

• Making gokr-rsync chroot (and/or Linux mount namespaces when available?) into the destination directory to reduce chances of accidental file system manipulation in case of bugs.

This project accepts contributions as time permits to merge them (best effort).

How do I know this project won’t eat my data?

This rsync implementation is very fresh. It was started in 2021 and doesn’t have many users yet.

With that warning out of the way, the rsync protocol uses MD4 checksums over file contents, so at least your file contents should never be able to be corrupted.

There is enough other functionality (delta transfers, file metadata, special files like symlinks or devices, directory structures, etc.) in the rsync protocol that provides opportunities for bugs to hide.

I recommend you carefully check that your transfers work, and please do report any issues you run into!

Existing rsync implementation survey

Language	URL	Note	Max Protocol	Server mode?
C	RsyncProject/rsync (formerly WayneD/rsync)	original “tridge” implementation; I found older versions easier to study	31	✔ yes
C	kristapsdz/openrsync	OpenBSD, good docs	27	✔ yes
Go	gokrazy/rsync	→ you are here ←	27	✔ yes 🎉
Go	jbreiding/rsync-go	rsync algorithm		❌ no
Go	kaiakz/rsync-os	only client/receiver	27	❌ no
Go	knight42	proxy		❌ no
Go	c4milo/gsync			❌ no
Java	APNIC-net/repositoryd	archived		✔ yes
Java	JohannesBuchner/Jarsync	archived, internet draft RFC “The rsync Network Protocol”		✔ yes
Java	perlundq/yajsync			✔ yes
C++	gilbertchen/acrosync-library	commercial		❌ no
Rust	sourcefrog/rsyn	client, “rsyn is rsync with no c”	27	❌ no

Getting started

To serve the current directory via rsync on localhost:8730, use:

go install github.com/gokrazy/rsync/cmd/gokr-rsync@latest
gokr-rsync --daemon --gokr.listen=localhost:8730 --gokr.modulemap=pwd=$PWD

You can then copy the contents of the current directory with clients such as rsync(1):

% rsync -v --archive --port 8730 rsync://localhost/pwd/ quine
receiving file list ... done
created directory quine
./
.git/
[…]
.github/workflows/main.yml
LICENSE
Makefile
README.md
cmd/gokr-rsyncd/rsyncd.go
doc.go
go.mod
go.sum
internal/rsyncd/connection.go
internal/rsyncd/rsyncd.go
interop_test.go

sent 1,234 bytes  received 5,678 bytes  13,824.00 bytes/sec
total size is 666  speedup is 0.10

…or openrsync(1), shown doing a differential update:

% openrsync -v --archive --port 8730 rsync://localhost/pwd/ quine
socket.c:109: warning: connect refused: ::1, localhost
Transfer starting: 369 files
.git/index (1.1 KB, 100.0% downloaded)
Transfer complete: 5.5 KB sent, 1.2 KB read, 666 B file size

Usage / Setup

setup	encrypted	authenticated	private files?	privileges	protocol version	config required
1. rsync daemon protocol (TCP port 873)	❌ no	⚠ rsync (insecure)	❌ only world-readable	✔ dropped + namespace	✔ negotiated	config required
2. anon SSH (daemon)	✔ yes	✔ rsync	❌ only world-readable	✔ dropped + namespace	✔ negotiated	config required
3. SSH (command)	✔ yes	✔ SSH	✔ yes	⚠ full user	⚠ assumed	no config
4. SSH (daemon)	✔ yes	✔ SSH (+ rsync)	✔ yes	⚠ full user	✔ negotiated	~/.config/gokr-rsyncd.toml required

Regarding protocol version “assumed”: the flags to send over the network are computed before starting SSH and hence the remote rsync process. You might need to specify --protocol=27 explicitly on the client. Once the connection is established, both sides do negotiate the protocol, though.

Setup 1: rsync daemon protocol (TCP port 873)

Serving rsync daemon protocol on TCP port 873 is only safe where the network layer ensures trusted communication, e.g. in a local network (LAN), or when using Tailscale or similar. In untrusted networks, attackers can eavesdrop on file transfers and possibly even modify file contents.

Prefer setup 2 instead.

Example:

• Server: gokr-rsync --daemon --gokr.modulemap=module=/srv/rsync-module
• Client: rsync rsync://webserver/module/path

Setup 2: anon SSH (daemon)

This setup is well suited for serving world-readable files without authentication.

Example:

• Server: gokr-rsync --daemon --gokr.modulemap=module=/srv/rsync-module --gokr.anonssh_listen=:22873
• Client: rsync -e ssh rsync://webserver/module/path

Setup 3: SSH (command)

This setup is well suited for interactive one-off transfers or regular backups, and uses SSH for both encryption and authentication.

Note that because gokr-rsync is invoked with user privileges (not root privileges), it cannot do namespacing and hence retains more privileges. When serving public data, it is generally preferable to use setup 2 instead.

Note that rsync(1) assumes the server process understands all flags that it sends, i.e. is running the same version on client and server, or at least a compatible-enough version. You can either specify --protocol=27 on the client, or use setup 4, which negotiates the protocol version, side-stepping possible compatibility gaps between rsync clients and gokr-rsync.

Example:

• Server will be started via SSH
• Client: rsync --rsync-path=gokr-rsync webserver:path

Setup 4: SSH (daemon)

This setup is more reliable than setup 3 because the rsync protocol version will be negotiated between client and server. This setup is slightly inconvenient because it requires a config file to be present on the server in ~/.config/gokr-rsyncd.toml.

Note that this mode of operation is only implemented by the original “trigde” rsync, not in openrsync. Apple started shipping openrsync with macOS 15 Sequoia, so you might need to explicitly start /usr/libexec/rsync/rsync.samba on Macs.

Example:

• Server will be started via SSH
• Client: rsync -e ssh --rsync-path=gokr-rsync rsync://webserver/module/path

Limitations

Bandwidth

In my tests, gokr-rsync can easily transfer data at > 6 Gbit/s. The current bottleneck is the MD4 algorithm itself (not sure whether in the “tridge” rsync client, or in gokr-rsync). Implementing support for more recent protocol versions would help here, as these include hash algorithm negotiation with more recent choices.

Protocol related limitations

• xattrs (including acls) was introduced in rsync protocol 30, so is currently not supported.

Supported environments and privilege dropping

Supported environments:

1. systemd (Linux)
2. Docker (Linux)
3. privileged Linux
4. privileged non-Linux

In all environments, the default instructions will take care that:

• (On Linux only) Only configured rsync modules from the host file system are mounted read-only into a Linux mount namespace for gokr-rsync, to guard against data modification and data exfiltration.
• gokr-rsync is running without privileges, as user nobody, to limit the scope of what an attacker can do when exploiting a vulnerability.

Known gaps:

• gokr-rsync does not guard against denial of service attacks, i.e. consuming too many resources (connections, bandwidth, CPU, …).
  • See also Per-IP rate limiting with iptables.

systemd (unprivileged)

We provide a gokr-rsyncd.socket and gokr-rsyncd.service file for systemd. These files enables most of systemd’s security features. You can check by running systemd-analyze security gokr-rsyncd.service, which should result in an exposure level of “0.2 SAFE” as of systemd 249 (September 2021).

First, configure your server flags by creating a systemd service override file:

systemctl edit gokr-rsyncd.service

In the opened editor, change the file to:

[Service]
ExecStart=
ExecStart=/usr/bin/gokr-rsync --daemon --gokr.modulemap=pwd=/etc/tmpfiles.d

Close the editor and install the service using:

systemctl enable --now gokr-rsyncd.socket

Additional hardening recommendations:

• Restrict which IP addresses are allowed to connect to your rsync server, for example:
  • using iptables or nftables on your host system
  • using gokr-rsync’s built-in IP allow/deny mechanism (once implemented)
  • using systemd’s IPAddressDeny and IPAddressAllow in gokr-rsyncd.socket
• To reduce the impact of Denial Of Service attacks, you can restrict resources with systemd, see Managing Resources.
• To hide system directories not relevant to any rsync module, use systemd’s TemporaryFileSystem= and BindReadOnlyPaths= directives as described in Use TemporaryFileSystem to hide files or directories from systemd services. Note that you may need to disable ProtectSystem=strict due to a bug.

Docker (unprivileged)

We provide a Dockerfile for gokr-rsyncd.

docker run \
  --read-only \
  -p 127.0.0.1:8730:8730 \
  -v /etc/tmpfiles.d:/srv/rsync:ro \
  stapelberg/gokrazy-rsync:latest \
    --gokr.modulemap=pwd=/srv/rsync

Additional hardening recommendations:

• Restrict which IP addresses are allowed to connect to your rsync server, for example:
  • using iptables or nftables on your host system
  • using gokr-rsync’s built-in IP allow/deny mechanism (once implemented)
    • Be sure to set up Docker such that the remote IPv4 or IPv6 address is available inside the container, see https://michael.stapelberg.ch/posts/2018-12-12-docker-ipv6/

privileged Linux (including gokrazy.org)

When started as root on Linux, gokr-rsync will create a Linux mount namespace, mount all configured rsync modules read-only into the namespace, then change into the namespace using chroot(2) and drop privileges using setuid(2).

Tip: you can verify which file system objects the daemon process can see by using ls -l /proc/$(pidof gokr-rsync)/root/.

Additional hardening recommendations:

• Restrict which IP addresses are allowed to connect to your rsync server, for example:
  • using iptables or nftables on your host system
  • using gokr-rsync’s built-in IP allow/deny mechanism (once implemented)

privileged non-Linux (e.g. Mac)

When started as root on non-Linux (e.g. Mac), gokr-rsync will drop privileges using setuid(2).

unprivileged with write permission (e.g. from a shell)

To prevent accidental misconfiguration, gokr-rsync refuses to start when it detects that it has write permission in any configured rsync module.

Documentation

Overview

Package rsync contains a native Go rsync implementation.

The only component currently is gokr-rsyncd, a read-only rsync daemon sender-only Go implementation of rsyncd. rsync daemon is a custom (un-standardized) network protocol, running on port 873 by default.

Index

• Constants
• type Logger
• type SumBuf
• type SumHead
  • func (sh *SumHead) ReadFrom(c *rsyncwire.Conn) error
  • func (sh *SumHead) WriteTo(c *rsyncwire.Conn) error

Constants

const (
	XMIT_TOP_DIR             = (1 << 0)
	XMIT_SAME_MODE           = (1 << 1)
	XMIT_EXTENDED_FLAGS      = (1 << 2)
	XMIT_SAME_RDEV_pre28     = XMIT_EXTENDED_FLAGS /* Only in protocols < 28 */
	XMIT_SAME_UID            = (1 << 3)
	XMIT_SAME_GID            = (1 << 4)
	XMIT_SAME_NAME           = (1 << 5)
	XMIT_LONG_NAME           = (1 << 6)
	XMIT_SAME_TIME           = (1 << 7)
	XMIT_SAME_RDEV_MAJOR     = (1 << 8)
	XMIT_HAS_IDEV_DATA       = (1 << 9)
	XMIT_SAME_DEV            = (1 << 10)
	XMIT_RDEV_MINOR_IS_SMALL = (1 << 11)
)

rsync.h

const (
	S_IFMT   = 0o0170000 // bits determining the file type
	S_IFDIR  = 0o0040000 // Directory
	S_IFCHR  = 0o0020000 // Character device
	S_IFBLK  = 0o0060000 // Block device
	S_IFREG  = 0o0100000 // Regular file
	S_IFIFO  = 0o0010000 // FIFO
	S_IFLNK  = 0o0120000 // Symbolic link
	S_IFSOCK = 0o0140000 // Socket
)

as per /usr/include/bits/stat.h:

const ProtocolVersion = 27

ProtocolVersion defines the currently implemented rsync protocol version. Protocol version 27 seems to be the safest bet for wide compatibility: version 27 was introduced by rsync 2.6.0 (released 2004), and is supported by openrsync and rsyn.

Types

type Logger

type Logger = log.Logger

Logger is an interface that allows specifying your own logger. By default, the Go log package is used, which prints to stderr.

type SumBuf

type SumBuf struct {
	Offset int64
	Len    int64
	Index  int32
	Sum1   uint32
	Sum2   [16]byte
}

rsync/rsync.h:struct sum_buf

type SumHead

type SumHead struct {
	// “number of blocks” (openrsync)
	// “how many chunks” (rsync)
	ChecksumCount int32

	// “block length in the file” (openrsync)
	// maximum (1 << 29) for older rsync, (1 << 17) for newer
	BlockLength int32

	// “long checksum length” (openrsync)
	ChecksumLength int32

	// “terminal (remainder) block length” (openrsync)
	// RemainderLength is flength % BlockLength
	RemainderLength int32

	Sums []SumBuf
}

TODO: remove connection.go:sumHead in favor of this type

func (*SumHead) ReadFrom

func (sh *SumHead) ReadFrom(c *rsyncwire.Conn) error

func (*SumHead) WriteTo

func (sh *SumHead) WriteTo(c *rsyncwire.Conn) error