Bytecode Alliance: One year update
bytecodealliance.org
bytecodealliance.org
Most platform apis let you just do stuff. Open /home/user/my-image.png. Ok.
The capabilities model re-orientes operations like this. The app starts with some kind of handle to a starting directory. Maybe it's to /, the root, maybe it's to /home/user, the user directory. Whoever has a reference to that directory handle can use it to open files / other directories inside that directory, but can not go upwards!
So if you have a file-saving middleware, you can be sure, whatever libraries that middleeware uses, that it will never write to anything other than the directory you give it.
This is a subtle & small change, with massive impact. Most OS have a parallel system-call that works similarly, openat(2), which is built around this idea of directory handles.
Capabilities are a more generic, core idea.
In this particular example, the capability to see a directory & it's children is what's getting passed around & refined (starting with the top level directory the app has, then perhaps winnowing down the capability to pass to the upload middleware).
But capabilities can represent other things too. The program might get capabilities to open a socket, to listen on a socket, to send or receive data on a socket. Those capabilities might similarly be refined & passed around to libraries.
The broader model of capabilities is that you can only do whatever you are passed, only exercise whatever capabilities you get. You have no other way to talk to the platform, no platform api, other than the set of capabilities you get from your caller. So capabilities keep getting winnowed down, shrunk, to the right size, where we expect, for the file-upload middleware, that it can read some bytes off some sockets it gets passed, and where it can write files, into a certain directory, but having no other capabilities, it can not do any more than that.
While filesystem sandboxing was core to my example, it's just one demonstration of what a capability might be. The underlying model for how that sandboxing is implemented is "capability based". There's a pretty long history for capability based systems. Not a simple example, but the "E language" is a pretty well known example from 1997 that tried to push capabilities into an interesting distributed frontier: http://www.erights.org/
I remember reading about a vulnerability in the unix utility "strings". The code is incredibly simple at first glance, but it had a dependency to detect the filetype, and that dependency was not safe on untrusted input.
At that moment I realized that unix security was fundamentally flawed. A utility that does nothing but read its input and write the output shouldn't have permission to do anything else.
- tail recursion elimination. There's a draft proposal but only one engine implements it at the moment which is blocking further progress.
- RISC-V backend for Cranelift (TBH, I'm sure someone will do it eventually).