FreeBSD 9.0, released in January 2012, has experimental Capsicum support in the kernel, disabled by default. In FreeBSD 10, Capsicum will be enabled by default.

But unless code uses it, we get no benefit. So far, very little code uses Capsicum, mostly just experiments we did for our paper. I figured it was time to start changing that. Today, I’ll describe my first venture – sandboxing bzip2. I chose bzip2 partly because Ilya Bakulin had already done some of the work for me, but mostly because a common failure mode in modern software is mistakes made in complicated bit twiddling code, such as decompressors and ASN.1 decoders.

These can often lead to buffer overflows or integer over/underflows – and these often lead to remote code execution. Which is bad. bzip2 is no stranger to this problem: CVE-2010-0405 describes an integer overflow that could lead to remote code execution. The question is: would Capsicum have helped – and if it would, how practical is it to convert bzip2 to use Capsicum?

The answers are, respectively, “yes” and “fairly practical”.

First of all, how does Capsicum mitigate this problem? The obvious way to defend a decompressor is to run the decompression engine in a separate process with no privilege beyond that needed to get its job done – which is the ability to read the input and write the output. In Capsicum, this is easy to achieve: once the appropriate files are open, fork the process and enter capability mode in the child. Discard all permissions except the ability to read the input and write the output (in Capsicum, this means close all other file descriptors and limit those two to read and write), and then go ahead and decompress. Should there be a bug in the decompressor, what does the attacker get? Well, pretty much what he had already: the ability to read the input file (he supplied it, so no news there!) and the ability to write arbitrary content to the output file (he already had that, since he could have chosen arbitrary input and compressed it). He also gets to burn CPU and consume memory. But that’s it – no access to your files, the network, any other running process, or anything else interesting.

I think that’s pretty neat.

But how hard is it to do? I answer that question in a series of diffs on GitHub, showing a step-by-step transformation of bzip2 into the desired form. I used a technique I like to call error-driven development; the idea is you attempt to make changes that will cause compilation to fail until you have completely accomplished your goal. This is a useful way to reassure yourself that you have made all necessary updates and there’s nothing hiding away you didn’t take care of. If you follow along by building the various stages, you’ll see how it works.

It turns out that in bzip2 this matters – it isn’t very beautifully written, and the code that looks like it might cleanly just take an input file and an output file and do the work in isolation, actually interacts with the rest of the code through various function calls and globals. This causes a problem: once you’ve forked, those globals and functions are now in the wrong process (i.e. the child) and so it is necessary to use RPC to bridge any such things back to the parent process. Error-driven development assures us that we have caught and dealt with all such cases.

So how did this work out in practice? Firstly, it turns out we have to give the compressor a little more privilege: it writes to stderr if there are problems, so we need to also grant write on stderr (note that we could constrain what it writes with a bit more effort). The callbacks we have to provide do not, I think, let it do anything interesting: cause the program to exit, make the output file’s permissions match the input file’s, and remove the input or output files (ok, removing the input file is slightly interesting – but note that bzip2 does this anyway).

Secondly, because we have not yet decided on an RPC mechanism, this particular conversion involves quite a bit of boilerplate: wrapping and unwrapping arguments for RPCs, wiring them up and all that, all of which would be vastly reduced by a proper RPC generator. Try not to let it put you off 

Finally, the system has (at least) one global, errno. I did not deal with that so far, which means some errors will report the wrong error – but it is not particularly hard to do so.

So, on to the diffs. This is something of an experimental way to present a piece of development, so I’d be interested in feedback. Here they are, in order:

• Step 1: move functions to be wrapped.
• Step 2: add header for wrapped functions.
• Step 3: wrap a function requiring authority.
• Step 4: move functions that do not require authority.
• Step 5: rename and invoke wrapped functions.
• Step 6: build other files.
• Step 7: wrap a global.
• Step 8: fix similar problems.
• Step 9: wrap a function in the parent.
• Step 10: wrap a 2-way global.
• Step 11: continue to apply the methodology.
• Step 12: get back on track with error-driven development.
• Step 13: define unwrappers.

And there you are: bzip2 is now rendered safe from decompressor exploits, and it was only a few hours work. As we refine the support infrastructure, it will be even less work.