Hacker News Reader

Comment by kiwicopple | Hacker News Reader

for those who want to know the difference between "trusted" and "untrusted":

https://tcdi.github.io/plrust/trusted-untrusted.html

_Normally, PL/Rust is installed as a "trusted" programming language named plrust. In this setup, certain Rust and pgx operations are disabled to preserve security. In general, the operations that are restricted are those that interact with the environment. This includes file handle operations, require, and use (for external modules). There is no way to access internals of the database server process or to gain OS-level access with the permissions of the server process, as a C function can do. Thus, any unprivileged database user can be permitted to use this language._

Languages like pl_python are "untrusted" and give too much access to the file system, which is why cloud providers never support them on their platforms

View on HN

cannot rust functions do anything a c function can?

only when using `unsafe` which this forbids

I don't know much about rust, so it seemed odd to me that just forbidding unsafe would make for a reasonable sandbox. At least for the postgres concept of "trusted language", that's supposed to mean things like closing off access to the filesystem. Not just typical file io calls either, but more obscure stuff like sendfile().

I do see that rust access to sendfile() would be via a syscall, which is in the unsafe category...so perhaps that's not the best example.

But it does make me curious how comprehensive a sandbox PL/Rust is providing, beyond just forbidding unsafe.

pl/rust is its own target platform, so they provide their own standard library*. The IO stuff all panics, for example. (In Rust a panic should be used to indicate invariant failure. pl/rust catches panics and converts them to postgres errors).

In Rust you'd normally be able to link c code, but calling c requires unsafe because you have to manually ensure the c code upholds any relevant Rust invariants.

> But it does make me curious how comprehensive a sandbox PL/Rust is providing, beyond just forbidding unsafe.

They also hook the compiler and try and detect shenanigans. It's not perfect, but it's pretty thought out.

*Technically the Rust standard library builds on top of a lower-level io module, which is all you have to replace.

More than just forbidding unsafe, but not enough to make this secure against competent adversaries by their own admission. They argue postgres itself isn't secure against competent adversaries so this doesn't matter too much.

https://tcdi.github.io/plrust/plrust.html#what-about-rust-co...

Rust keeps a list of soundness bugs via a tag on github - they're pretty common:

https://github.com/rust-lang/rust/issues?q=is%3Aopen+is%3Ais...

Read the article, they specify that they use a compiler target that restricts system access.

As I read it, it seemed that the unsafe prohibition was more about the safety of the code and not the security of the system.

And there is also an alternative implementation to Rust std that blocks, among others, filesystem access via std::fs https://github.com/tcdi/postgrestd

This implementation blocks file system access and is thus not vulnerable, but note that in Rust in general you can actually violate safety on some platforms without any `unsafe` by modifying magic files like /proc/self/mem. This is a known issue but considered unfixable (because the technical solution of marking opening a file as unsafe would cause far more trouble than it could ever hope to solve).

It's more of an interesting fact than an issue to be fixed.

Normally yes, but it looks like the trusted PL/Rust being discussed here is limited to some subset of Rust. They specifically note that `unsafe` code is not allowed, which means you can't (for example) implement your own syscalls or construct a pointer into postgres internals memory.

However, they make it clear that this is not intended to be your only defence against an attacker:

> Note that this is done on a best-effort basis, and does not provide a strong level of security — it's not a sandbox, and as such, it's likely that a skilled hostile attacker who is sufficiently motivated could find ways around it

I don't think it's really any defence at all against an attacker. I don't think Rust's `unsafe` was ever intended to be a security boundary.

Surely all a "sufficiently motivated" attacker would need to do is peruse the unsound bugs on GitHub?

https://github.com/rust-lang/rust/issues?q=is%3Aopen+is%3Ais...

Those aren't considered to be security issues. Makes me wonder what the point of banning `unsafe` is at all. You're going to need some other system anyway...

The trust is not just banning unsafe, it is using a limited std:

> The "trusted" version of PL/Rust uses a unique fork of Rust's std entitled postgrestd when compiling LANGUAGE plrust user functions.

https://github.com/tcdi/postgrestd

That doesn't matter. If you can use `unsafe` you can make syscalls directly.

A large number of unsoundness bugs only work if you have access to the stdlib, because they're flaws in stdlib types and functions that use `unsafe` internally, and are supposed to present a safe interface around it.

If you a) don't have access to unsafe, and b) don't have access to the stdlib that lets you do powerful things without unsafe, then you're very limited in what you can do.

https://smallcultfollowing.com/babysteps/blog/2016/10/02/obs... discusses this further. Conceptually, you can think of "entirely Safe Rust" to be a very limited language, which you then progressively add "capabilites" to by exposing safe interfaces implemented with unsafe code. For example, Vec and Box (which require unsafe) grant safe code the ability to do heap allocations.

It's true that this is not designed as a security boundary. As I note in my comment above, the PL/Rust devs also make that clear. That doesn't mean it has no value as part of a defence in depth strategy.

The rustc driver for trusted PL/Rust prevents using the subset of the Rust language required to trigger those issues. Most of them are things that would have a hard time traversing the Postgres procedure call boundary, anyways, in a legitimate use-case, so this isn't expected to meaningfully affect actual user code.

So you have to give them your rust code to compile on your behalf then?

Reply on news.ycombinator.com