Summary

Varnish Cache with certain configurations is vulnerable to a denial
of service attack.

Three lines of VCL code solves the problem.

This issue was discovered by Ilia Sharov, Yandex.

This has been assigned CVE-2013-4484.

Details

If Varnish receives a certain illegal request, and the subroutine
'vcl_error{}' restarts the request, the varnishd worker process
will crash with an assert.

The varnishd management process will restart the worker process,
but there will be a brief interruption of service and the cache
will be emptied, causing more traffic to go to the backend.

We are releasing this advisory because restarting from vcl_error{}
is both fairly common and documented.

This is purely a denial of service vulnerability, there is no risk
of privilege escalation.

Proof of concept

Given a VCL with the effect of:

sub vcl_error {
    return(restart);
}

and a malformed HTTP request like:

GET<SP><SP><SP><CR><NL>
Host:<SP>foo<CR><NL>
<CR><NL>

Varnish will assert and restart the child process.

(The precise number of spaces after GET is not magic.)

Cause

A malformed request should never reach the VCL processing in Varnish
in the first place, but for historical reasons we used vcl_error{}
to deliver the error-response for some malformed requests.

In future versions of Varnish, (4.x, 3.0.5) a standard summary 400
response will be returned for all requests which cannot be parsed
correctly, without VCL involvement.

Workaround

Insert this at the top of your VCL file:

sub vcl_error {
    if (obj.status == 400 || obj.status == 413) {
        return(deliver);
    }
}

Or add this test at the top of your existing vcl_error{}.

Versions affected

At least 2.0.x, 2.1.x, 3.0.x, possibly also older versions.

– Poul-Henning Kamp | UNIX since Zilog Zeus 3.20 [email protected] | TCP/IP since RFC 956 FreeBSD committer | BSD since 4.3-tahoe Never attribute to malice what can adequately be explained by incompetence.