CloudABI - Easily develop sandboxed apps for UNIX

This article is developed from a talk by Ed Schouten at FOSDEM 2017. This article piqued my interest due to my dissertation being on a very similar subject, and with the focus on the Cloud.

Project Rationale

CloudABI is a project born out of the need to harden applications such that exploits are unable to cause any undue access or damage to the host machine. It aims to make the process of sandboxing these applications much easier, such that developers don't have to jump through hoops in order to get a level of security.

There currently are a few security frameworks around that provide such functionality, such as AppArmor, SELinux and POSIX 1003.1e capabilities which run on the OS around the application. However, these are usually built with the mindset that the default security model - Discretionary Access Control ("I, firefox, am running as the user jamie, I have have access to ~/.ssh/* because I should be able to access any files that jamie can, because it says I do") - isn't best. Instead, they provide an extra level of security in the kernel, that provides extra security to prevent any access through Mandatory Access Control ("even though, I, firefox, am running as the user jamie, I don't have access to ~/.ssh/* because I should never need access to the public keys").

The issue with these frameworks is that they provide the wrapping external to the application. That is, the application itself isn't developed with this in mind. Because of this, the rulesets required to make it function within the framework are created by observing the system in use. However, this can lead to false positives, and assumptions that certain steps in the application are well-formed, when in fact they may not be.

Therefore, as Ed goes on to discuss, applications should be built with this sandboxing inherent, such that they can ensure they will always perform valid actions. Additionally, the application developer should always know what their code will be doing, and how it should perform. Ed continues to how the problem should be fixed at the root, rather than at the top - if the application itself is defined with the well-formed routes, and the sandbox it runs in works as expected, then the issue of hardening should be solved.

One such tool that provides an opt-in sandboxing and capabilities-based framework is Capsicum, an Object Oriented Operating System API which allows developers to create their applications with these sandboxes in mind. Capsicum works by providing file descriptors that have fine grained controls, such that an application will not access any arbitrary resources. Because these file descriptors are simply a handle to some resource, they don't provide implicit read-write access, much like that of a Mandatory Access Control system, means that even if a process is running as root, it still doesn't have the full system access as it has to work within the file descriptors provided to it. This means that a process would only be able to open a socket at /tmp/tmux.1000, instead of /tmp/tmux-socket or /tmp/tmux-x11 or similar.

However, there are a number of things that don't work out of the box - setting the timezone, using tzset, and even open('/dev/urandom'). These inconveniences, for programs that would expect such behaviour to just work brought the comment "sandboxing is stupid, and you shouldn't use it!" Ed went on to discuss how often, you spend far too long working out why the program isn't working, instead of working on more important tasks. And "even if it works, not necessarily working as intended"; the idea that just because the rules have been configured to make sure that the application works, doesn't mean it does the right things. For instance, the application could be malformed, and could be taking steps that it shouldn't be, such as overreaching the file access it's making.

So What Can CloudABI Do?

But this still doesn't make things optimal - what if we made it so we had unconditional sandboxing? And if any incompatible APIs were completely removed from the built application? Where anything that is compatible with the APIs, is implemented to work well with sandboxing? And what if that optimal system was enforced at build-time, not run-time? This is where we get to CloudABI. In a world where this is enforced at build-time, you would have to work against compiler errors, incrementally fixing your application against the static set of issues, until you finally had a build. This final version would be well-formed, and would not allow anything that is unsupported by the ABI.

CloudABI was built by taking the C functions in the POSIX spec, then adding common extensions following the Capsicum API, and removing anything that would break the API. This still works, however, and provides just under 60 syscalls, ensuring to remove syscalls that are prone to buffer overflows, and similar issues. This resulting set of syscalls provides an Operating System independent ABI for sandboxed applications; and additionally, one that is much more lean than a standard Operating System (for reference, Linux has ~300 syscalls).

CloudABI also provides a tool, cloudabi-run which provides an easy wrapper to run programs through, using the CloudABI YAML syntax. This makes it possible to define a C program, such as:

#include <argdata.h>
#include <program.h>
#include <stdio.h>
#include <stdlib.h>

void program_main(const argdata_t *ad) {
    argdata_seq_iterator_t it;
    argdata_seq_iterate(ad, &it);
    const argdata_t *value;
    while (argdata_seq_next(&it, &value)) {
        int fd;
        if (argdata_get_fd(value, &fd) == 0)
            dprintf(fd, "Hello, world\n");
    }
    exit(0);
}

With the following YAML:

%TAG	! tag:nuxi.nl,2015:cloudabi/
---
- !fd stdout
- !fd stdout
- !fd stdout

Which can be run as:

$ x86_64-unknown-cloudabi-cc	-o hello hello.c
$ cloudabi-run hello	< hello.yaml
Hello, world
Hello, world
Hello, world

This shows that it can be made very simple, and that CloudABI is quite an elegant, and simple solution. In addition, it helps ensure that you can't accidentally leak a file descriptor to a given application - they must be manually pushed to the application.

CloudABI has its own machine-generated API under the guise cloudlibc. This provides a sandboxing library, with its own library of functionality such as time.h to avoid any issues with the aforementioned tzset. cloudlibc itself has a number of prepackaged 'CloudABI Ports' such as libc++, Boost, Python3, json-c and basic Django support (sans templating support), to name a few. Additionally, unlike other projects which require their own package managers, CloudABI produces native packages, such that things can be installed easily, and without another set of dependencies.

A big perk of CloudABI is that it means that you will be able to test your applications much more easily, as it works almost as a "dependency injection system for UNIX" - each application's resources can be easily swapped out, with the underlying application not needing a recompile, or needing to even be aware of it, as long as they're able to work with them.

One caveat with using CloudABI is that it is not recommended to use with third party libraries - that is, don't arbitrarily add CloudABI enforcement on a third party library. This is because it will meant that you will need to "fight their code to make it work", as they will most likely have not designed their code with CloudABI or sandboxing in mind - otherwise they'd already be using it. Therefore, you will spend a large amount of time determining the inner workings of their product, and trying to ensure that the code can be worked around in order to make sure that the sandbox works. But as soon as they do an update, all your work could be for nothing, and you'd then have to go back to hacking around everything and making it work under the sandbox. However, if everyone adopts it, then there will be no issues - all libraries and pieces of software will find it possible to run on top of CloudABI.

CloudABI is available on current platforms - FreeBSD works out-of-the-box, Linux and NetBSD require patches to add full enforcement, and MacOS emulates but does not actually enforce it.

Written by Jamie Tanna's profile image Jamie Tanna on , and last updated on .

Content for this article is shared under the terms of the Creative Commons Attribution Non Commercial Share Alike 4.0 International, and code is shared under the Apache License 2.0.

#events #fosdem #cloud #open-source.

This post was filed under articles.

Has this content helped you? Please consider supporting me so I can continue to create content like this!

Related Posts

Other posts you may be interested in:

Interactions with this post

Interactions with this post

Below you can find the interactions that this page has had using WebMention.

Have you written a response to this post? Let me know the URL:

Do you not have a website set up with WebMention capabilities? You can use Comment Parade.