Executes a program and converts IP to Unix domain sockets at runtime via LD_PRELOAD (ld.so(8)) based on a list of rules, either given via short command line options (see Rule specification) or via a file containing a list of rules separated via newline. The first matching rule causes ip2unix to replace the current IP socket with a Unix domain socket based on the options given. For example if a path is specified, the Unix domain socket will bind or listen to the file given.
A lot of programs are designed to only work with IP sockets, however very few of them allow to communicate via Unix domain sockets. Unix domain sockets usually are just files, so standard Unix file permissions apply to them.
IP sockets also have the disadvantage that other programs on the same host are able to connect to them, unless you use complicated netfilter rules or network namespaces.
So if you either have a multi-user system or just want to separate privileges, Unix domain sockets are a good way to achieve that.
Another very common use case in nowadays systems is when you’re using systemd and want to use socket units for services that don’t support socket activation. Apart from getting rid of the necessity to specify explicit dependencies, this is also very useful for privilege separation, since a lot of services can be run in a separate network namespace.
The systemd use case is also useful even when not using Unix domain sockets in socket units, since it allows to add IP-based socket activation to services that don’t support it.
Let’s say you have a small HTTP server you want to make available behind a HTTP reverse proxy.
$ ip2unix -r path=/run/my-http-server.socket my-http-serverThis will simply convert all IP sockets to the Unix domain socket available at
/run/my-http-server.socket. If you use a web server like
nginx, you can use the following directive to connect to
that socket:
proxy_pass http://unix:/run/my-http-server.socket;More examples can be found further below in section Examples.
A short summary of all the options is available via ip2unix --help or
man ip2unix if you want to see all the details and options available.
-
AsciiDoc or Asciidoctor for generating the manpage. The former is recommended as it generates a better manpage and also provides validation.
-
pytest for running automated tests.
-
systemd-socket-activate helper to run test cases specific to systemd socket activation support.
The source code can be fetched via Git using the following command:
$ git clone https://github.com/nixcloud/ip2unix.gitYou will get an ip2unix directory inside your current working directory. All
of the following steps are to be performed inside this ip2unix directory.
This is the easiest and recommended way to compile it from source and it should work on any distribution.
$ curl https://nixos.org/nix/install | shIn order to build ip2unix issue the following command from the top of the source tree:
$ nix-buildThis takes care of fetching the dependencies, building and running the test
suite. The resulting command can now be found in result/bin/ip2unix.
If you want to add the package to your user environment, you can install it using the command:
$ nix-env -f . -iTo install the required dependencies:
$ sudo apt install meson g++If you want to have the manpage:
$ sudo apt install asciidoctorIn case you want to run the test suite, pytest is required:
$ sudo apt install python3-pytestTo install the required dependencies:
$ sudo yum install meson gcc-c++If you want to have the manpage:
$ sudo yum install asciidoctorIf you want to run the test suite:
$ sudo yum install python3-pytest$ meson buildIf you want to specify a different compiler executable, eg. g++-7:
$ CXX=g++-7 meson buildCompile:
$ ninja -C buildThe executable is then placed in build/ip2unix, so to show the usage:
$ build/ip2unix --helpTo install ip2unix, run the following command:
$ ninja -C build installBy default, this will install ip2unix in /usr/local/bin/ip2unix.
Arguments specified via -r contain a comma-separated list of either flags or
options. If a value contains a comma (,), it has to be escaped using a
backslash (\) character. If you want to have a verbatim backslash character
just use two consecutive backslashes instead.
The following flags are available:
- in | out
-
Whether this rule applies to a server-side socket (
in), a client-side socket (out) or both if neitherinnoroutis specified. - tcp | udp
-
Specifies the IP type, which currently is either
tcpfor TCP sockets,udpfor UDP sockets or if it is not defined it matches both UDP and TCP sockets. - systemd[=FD_NAME]
-
Use the socket passed along via file descriptor by systemd instead of creating one.
An optional file descriptor name (FD_NAME) can be specified to distinguish between several socket units. This corresponds to the FileDescriptorName systemd socket option.
- reject[=ERRNO]
-
Reject calls to connect and bind with
EACCESby default or the ERRNO specified either via name or as an integer. - blackhole
-
When binding the socket, use a temporary file system path and unlink it shortly after the bind. This is a way to deactivate a specific socket without the application noticing.
- ignore
-
Prevents a socket from being converted to a Unix domain socket if this is set. This is useful to exempt specific sockets from being matched when another rule matches a broad scope.
These options are available:
- addr[ess]=ADDRESS
-
The IP address to match, which can be either an IPv4 or an IPv6 address.
- port=PORT[-PORT_END]
-
UDP or TCP port number which for outgoing connections specifies the target port and for incomping connections the port that the socket is bound to.
If a range is specified by separating two port numbers via
-, the given range is matched instead of just a single port. The range is inclusive, so if2000-3000is specified, both port 2000 and port 3000 are matched as well. - path=SOCKET_PATH
-
The path to the socket file to either bind or connect to.
Placeholders are allowed here and are substituted accordingly:
%p port number
%a IP address or
unknown%t socket type (
tcp,udporunknown)%% verbatim
%
The following command spawns a small test web server listening on
/tmp/test.socket:
$ ip2unix -r in,path=/tmp/test.socket python3 -m http.server 8000This connects to the above test server listening on /tmp/test.socket and
should show a directory listing:
$ ip2unix -r out,path=/tmp/test.socket curl http://1.2.3.4/For example the following could be put into a file given by the -f command line argument:
out,port=53,ignore out,tcp,path=/run/some.socket in,addr=1.2.3.4,path=/run/another.socket in,port=80,address=abcd::1,blackhole in,port=80,reject=EADDRINUSE in,tcp,port=22,systemd=ssh
Each line corresponds to a single rule, that is processed in order of appearance and the above example would result in the following:
-
All outgoing connections to port 53 (no matter if it’s TCP or UDP) will not be converted into Unix domain sockets.
-
This rule will redirect all TCP connections except to port 53 (see above) to use the Unix domain socket at
/run/some.socket. -
Matches the socket that listens to any port on the IPv4 address
1.2.3.4and instead binds it to the Unix domain socket at/run/another.socket. -
The application may bind to the IPv6 address
abcd::1on port 80 but it will not receive any connections, because no socket path exists. -
Trying to bind to port 80 on addresses other than
abcd::1will result in anEADDRINUSEerror. -
Will prevent the TCP socket that would listen on port 22 to not listen at all and instead use the systemd-provided file descriptor named
sshfor operations like accept(2).
The same can be achieved solely using -r commandline arguments:
$ ip2unix -r out,port=53,ignore \
-r out,tcp,path=/run/some.socket \
-r in,addr=1.2.3.4,path=/run/another.socket \
-r in,port=80,address=abcd::1,blackhole \
-r in,port=80,reject=EADDRINUSE \
-r in,tcp,port=22,systemd=ssh-
The program uses LD_PRELOAD (ld.so(8)), so it will only work with programs that are dynamically linked against the C library. Using ip2unix on statically linked executables or on executables that don’t use the socket family functions of the C library (like Go programs) will not work at the moment.
-
If a client which is already using Unix datagram sockets sends packets via sendto or sendmsg to a socket provided by ip2unix without binding first, ip2unix is not able to identify the peer and will subsequently reject the packet. This is not the case when using ip2unix itself on the the client side and it also does not seem to be very common as the author so far did not find such an application in the wild.
However, if this really is an issue to you, the recommended workaround is either to use ip2unix to wrap the client (if it supports IP sockets) or fix the server to natively use Unix domain sockets.
5.1. Isn’t this functionality already covered by socat?
The socat tool has a very different purpose: It is essentially a way of connecting streams between different address types. Apart from a myriad of options, it supports quite a lot of address types and it’s really good at providing great flexibility to connect bidirectional streams.
However what it doesn’t do is change the behaviour of the target application, which is what ip2unix does.
For example, if you have an application that listens to TCP port 1234, you can
use socat to create a Unix domain socket listening on foo.sock and proxying
all requests to TCP port 1234:
$ socat UNIX-LISTEN:foo.sock,fork TCP:localhost:1234Here, the application will still listen to TCP port 1234, but we now have two additional sockets (Unix inbound and TCP/IP outbound) we need to take care of.
ip2unix on the other side redirects the C library calls of the application in question, so that TCP port 1234 will not be bound in the first place and instead the application directly binds to a Unix domain socket.
This not only allows for better privilege separation (because local users need file system access permissions to the socket file) but also involves less overhead since only one socket (the listening socket of the application itself) is used.
5.2. Yes, but can’t this still be done via socat and the iptables owner module?
Of course you could use iptables to only allow access to the user running socat. But again, this still needs additional sockets and also still doesn’t decrease the attack surface by a large margin (eg. there could be race conditions in loading iptables rules or simply human error specifying the rules).
Not binding to an IP socket in the first place however gets rid of that attack surface, since you can’t attack things that don’t exist.
- socket_wrapper
-
The goal is a different one here and its main use is testing. Instead of using rules, socket_wrapper turns all of the IP sockets into Unix sockets and uses a central directory to do the mapping.
Containing all Unix sockets into one directory has the nice effect that it is easy to map any address/port combination to Unix sockets. While this is way easier to implement than our approach it has the drawback that everything is contained and no IP communication is possible anymore.
Special thanks to the NLnet foundation for sponsoring the initial work on this project.
Copyright © 2018 aszlig. License LGPLv3: GNU LGPL version 3 only https://www.gnu.org/licenses/lgpl-3.0.html.
This is free software: you are free to change and redistribute it. There is NO WARRANTY, to the extent permitted by law.