Security barriers in Varnish

Security is a very important design driver in Varnish, more likely than not, if you find yourself thinking "Why did he do _that_ ? the answer has to do with security.

The Varnish security model is based on some very crude but easy to understand barriers between the various components:

              .-->- provides ->---------------------------------------.
              |                                          |            |
     (ADMIN)--+-->- runs ----->---.                      |            |
              |                   |                      |            |
              |-->- cli_req -->---|                      v            v
              '--<- cli_resp -<---|                     VCL        MODULE
                                  |                      |            |
     (OPER)                       |                      |reads       |
       |                          |                      |            |
       |runs                      |                      |            |
       |      .-<- create -<-.    |    .->- fork ->-.    v            |
       v      |->- check -->-|-- MGR --|            |-- VCC <- loads -|
      VSM     |-<- write --<-'    |    '-<- wait -<-'    |            |
     TOOLS    |                   |                      |            |
       ^      |     .-------------'                      |            |
       |      |     |                                    |writes      |
       |reads |     |->- fork ----->-.                   |            |
       |      |     |->- cli_req -->-|                   |            |
      VSM ----'     |-<- cli_resp -<-|                   v            |
       |            '-<- wait -----<-|                VCL.SO          |
       |                             |                   |            |
       |                             |                   |            |
       |---->----- inherit --->------|--<-- loads -------'            |
       |---->-----  reads ---->------|                                |
       '----<----- writes ----<------|--<-- loads --------------------'
                                     |
                                     |
                                     |
         .--->-- http_req --->--.    |    .-->-- http_req --->--.
(ANON) --|                      |-- CLD --|                     |-- (BACKEND)
         '---<-- http_resp --<--'         '--<-- http_resp --<--'

(ASCII-ART rules!)

The really Important Barrier

The central actor in Varnish is the Manager process, "MGR", which is the process the administrator "(ADMIN)" starts to get web-cache service.

Having been there myself, I do not subscribe to the "I feel cool and important when I get woken up at 3AM to restart a dead process" school of thought, in fact, I think that is a clear sign of mindless stupidity: If we cannot get a computer to restart a dead process, why do we even have them ?

The task of the Manager process is therefore not cache web content, but to make sure there always is a process which does that, the Child "CLD" process.

That is the major barrier in Varnish: All management happens in one process all actual movement of traffic happens in another, and the Manager process does not trust the Child process at all.

The Child process is in a the totally unprotected domain: Any computer on the InterNet "(ANON)" can connect to the Child process and ask for some web-object.

If John D. Criminal manages to exploit a security hole in Varnish, it is the Child process he subverts. If he carries out a DoS attack, it is the Child process he tries to fell.

Therefore the Manager starts the Child with as low priviledge as practically possible, and we close all filedescriptors it should not have access to and so on.

There are only three channels of communication back to the Manager process: An exit code, a CLI response or writing stuff into the shared memory file "VSM" used for statistics and logging, all of these are well defended by the Manager process.

The Admin/Oper Barrier

If you look at the top left corner of the diagram, you will see that Varnish operates with separate Administrator "(ADMIN)" and Operator "(OPER)" roles.

The Administrator does things, changes stuff etc. The Operator keeps an eye on things to make sure they are as they should be.

These days Operators are often scripts and data collection tools, and there is no reason to assume they are bugfree, so Varnish does not trust the Operator role, that is a pure one-way relationship.

(Trick: If the Child process us run under user "nobody", you can allow marginally trusted operations personel access to the "nobody" account (for instance using .ssh/authorized_keys2), and they will be able to kill the Child process, prompting the Manager process to restart it again with the same parameters and settings.)

The Administrator has the final say, and of course, the administrator can decide under which circumstances that authority will be shared.

Needless to say, if the system on which Varnish runs is not properly secured, the Administrator's monopoly of control will be compromised.

All the other barriers

There are more barriers, you can spot them by following the arrows in the diagram, but they are more sort of "technical" than "political" and generally try to guard against programming flaws as much as security compromise.

For instance the VCC compiler runs in a separate child process, to make sure that a memory leak or other flaw in the compiler does not accumulate trouble for the Manager process.

Hope this explanation helps understand why Varnish is not just a single process like all other server programs.

Poul-Henning, 2010-06-28