Security first

If you are the only person involved in running Varnish, or if all the people involved are trusted to the same degree, you can skip this chapter. We have protected Varnish as well as we can from anything which can come in through an HTTP socket.

If parts of your web infrastructure are outsourced or otherwise partitioned along administrative lines, you need to think about security.

Varnish provides four levels of authority, roughly related to how and where the command comes into Varnish:

  • the command line arguments,
  • the CLI interface,
  • VCL programs, and
  • HTTP requests.

Command line arguments

The top level security decisions is decided and defined when starting Varnish in the form of command line arguments, we use this strategy in order to make them invulnerable to subsequent manipulation.

The important decisions to make are:

  1. Who should have access to the Command Line Interface?
  2. Which parameters can they change?
  3. Will inline-C code be allowed?
  4. If/how VMODs will be restricted?

CLI interface access

The command line interface can be accessed three ways.

Varnishd can be told til listen and offer CLI connections on a TCP socket. You can bind the socket to pretty much anything the kernel will accept:

-T 127.0.0.1:631
-T localhost:9999
-T 192.168.1.1:34
-T '[fe80::1]:8082'

The default is -T localhost:0 which will pick a random port number, which varnishadm(8) can learn in the shared memory.

By using a "localhost" address, you restrict CLI access to the local machine.

You can also bind the CLI port to an IP number reachable across the net, and let other machines connect directly.

This gives you no secrecy, ie, the CLI commands will go across the network as ASCII text with no encryption, but the -S/PSK authentication requires the remote end to know the shared secret.

Alternatively you can bind the CLI port to a 'localhost' address, and give remote users access via a secure connection to the local machine, using ssh/VPN or similar.

If you use ssh you can restrict which commands each user can execute to just varnishadm, or even to wrapper scripts around varnishadm, which only allow specific CLI commands.

It is also possible to configure varnishd for "reverse mode", using the '-M' argument. In that case varnishd will attempt to open a TCP connection to the specified address, and initiate a CLI connection to your central Varnish management facility.

The connection is also in this case without secrecy, but the remote end must still satisfy -S/PSK authentication.

Finally, if you run varnishd with the '-d' option, you get a CLI command on stdin/stdout, but since you started the process, it would be hard to prevent you getting CLI access, wouldn't it ?

CLI interface authentication

By default the CLI interface is protected with a simple, yet strong "Pre Shared Key" authentication method, which do not provide secrecy (ie: The CLI commands and responses are not encrypted).

The way -S/PSK works is really simple: During startup a file is created with a random content and the file is only accessible to the user who started varnishd (or the superuser).

To authenticate and use a CLI connection, you need to know the contents of that file, in order to answer the cryptographic challenge varnishd issues.

(XXX: xref to algo in refman) .. XXX:Dunno what this is? benc

varnishadm uses all of this to restrict access, it will only function, provided it can read the secret file.

If you want to allow other users, local or remote, to be able to access CLI connections, you must create your own secret file and make it possible for (only!) these users to read it.

A good way to create the secret file is:

dd if=/dev/random of=/etc/varnish_secret count=1

When you start varnishd, you specify the filename with '-S', and it goes without saying that the varnishd master process needs to be able to read the file too.

You can change the contents of the secret file while varnishd runs, it is read every time a CLI connection is authenticated.

On the local system, varnishadm can retrieve the filename from shared memory, but on remote systems, you need to give varnishadm a copy of the secret file, with the -S argument.

If you want to disable -S/PSK authentication, specify '-S' with an empty argument to varnishd:

varnishd [...] -S "" [...]

Parameters

Parameters can be set from the command line, and made "read-only" (using '-r') so they cannot subsequently be modified from the CLI interface.

Pretty much any parameter can be used to totally mess up your HTTP service, but a few can do more damage than others:

user and group
Access to local system via VCL
listen_address
Trojan other TCP sockets, like ssh
cc_command
Execute arbitrary programs
vcc_allow_inline_c
Allow inline C in VCL, which would any C code from VCL to be executed by Varnish.

Furthermore you may want to look at and lock down:

syslog_cli_traffic
Log all CLI commands to syslog(8), so you know what goes on.
vcc_unsafe_path
Restrict VCL/VMODS to vcl_dir and vmod_dir
vmod_dir
The directory where Varnish will will look for modules. This could potentially be used to load rouge modules into Varnish.

The CLI interface

The CLI interface in Varnish is very powerful, if you have access to the CLI interface, you can do almost anything to the Varnish process.

As described above, some of the damage can be limited by restricting certain parameters, but that will only protect the local filesystem, and operating system, it will not protect your HTTP service.

We do not currently have a way to restrict specific CLI commands to specific CLI connections. One way to get such an effect is to "wrap" all CLI access in pre-approved scripts which use varnishadm(1)

to submit the sanitized CLI commands, and restrict a remote user to only those scripts, for instance using sshd(8)'s configuration.

VCL programs

There are two "dangerous" mechanisms available in VCL code: VMODs and inline-C.

Both of these mechanisms allow execution of arbitrary code and will thus allow a person to get access to the machine, with the privileges of the child process.

If varnishd is started as root/superuser, we sandbox the child process, using whatever facilities are available on the operating system, but if varnishd is not started as root/superuser, this is not possible. No, don't ask me why you have to be superuser to lower the privilege of a child process...

Inline-C is disabled by default starting with Varnish version 4, so unless you enable it, you don't have to worry about it.

The parameters mentioned above can restrict the loading of VMODs to only be loaded from a designated directory, restricting VCL wranglers to a pre-approved subset of VMODs.

If you do that, we are confident that your local system cannot be compromised from VCL code.

HTTP requests

We have gone to great lengths to make Varnish resistant to anything coming in throught the socket where HTTP requests are received, and you should, generally speaking, not need to protect it any further.

The caveat is that since VCL is a programming language which lets you decide exactly what to do with HTTP requests, you can also decide to do stupid and potentially dangerous things with them, including opening youself up to various kinds of attacks and subversive activities.

If you have "administrative" HTTP requests, for instance PURGE requests, we strongly recommend that you restrict them to trusted IP numbers/nets using VCL's Access control lists (ACLs).