[dbaupp]

Thanks!

I'm not convinced that the statement in the paper translates into what you said: the key piece of that paragraph is "or seems to be". The Leakpocalypse problem was one piece of code (crossbeam's scoped threads API) was relying on an invariant that doesn't actually hold ("destructors will always run"). It was, fundamentally, a bug in the `unsafe` code in crossbeam, meaning it was incorrect for crossbeam to call its API safe: the fact that it took multiple libraries to trigger in that case means nothing, it just happens to be the circumstances under which the problem was noticed.

Of course, to be fair, no-one had thought about this destructor property before, just implicitly relied on it, and so it does demonstrate the necessity for better understanding of/tools for unsafe code, which is what projects like RustBelt are pushing towards.

To summarise, I still don't see how these two sentences are different:

> no unsafe behavior is observable by clients of the library

> [clients] without "unsafe" blocks cannot observe unsafe behavior in [the] library

Indeed, I don't think it makes sense to even attempt to prove that clients with unsafe code can't observe unsafe behaviour (which seems to be the only way for the second sentence to differ from the first). The typical framing is that the safe code can be arbitrarily bad and there'll still be no undefined behaviour, but arbitrary `unsafe` can do anything, including writing directly to another library's data structures, which of course can easily cause UB (e.g. replace a Vec's data pointer with a null one).

[klodolph]

> To summarise, I still don't see how these two sentences are different: ...

To "observe unsafe behavior" means I can write a program that does something safe, e.g., a data race or invalid memory access. It's possible to write library X and Y in such a way that I can observe unsafe behavior using both X and Y in my program, without putting "unsafe" blocks in my program. This is possible even if I can't do the same thing with either X or Y alone.

This is surprising, because it means that the naive definition of "safe interface" is not actually safe enough!

[dbaupp]

I'm still not understanding: other than the hardware/global state thing in another comment in this thread[1], what's a program that demonstrates this "composing safe interfaces is unsafe" property? The example in the paper is not one, it was a bug for crossbeam to mark its API safe.

[1]: I'm ignoring this case, because it's somewhat completely impossible to solve: there's no way Rust (or any language) can control this situation. And, there's a strong argument in my mind that this sort of scenario should have an `unsafe` constructor or something, to act as an assertion from the programmer that they're guaranteeing unique access to the resource.

[AstralStorm]

There is a way. You have to meticulously prove operations down to machine code to not have externally observable side effects.

You can weaken the condition by excluding, say, timing effects or cacheline effects. (Say hello to Spectre)

This means you get to prove bounded access and data race freedom on any piece of memory safe code touches. Likewise prove bounded access for all unsafe code and correct cpu flag and state handling.

It it's not as bad as it seems - you can use the machine code prover designed for seL4 as a good starting point.

[Rusky]

The definition of what you're allowed to do with `unsafe`, however nebulous its specifics may be at the moment, is that such a situation is a bug in one or both of those libraries, not their composition.

To put it another way, if you can't observe unsafety with X or Y alone, but you can with both together, then at least one of them has given you a new capability that you did not have before. Either that new capability is not truly safe, and thus the bug is providing that capability, or it exposes the other library relying on something not truly safe, and thus the bug is relying on that property.

The important point here is that, by definition, at least one of X or Y will have to change when such a situation is discovered, in order to preserve the property that composing safe interfaces is safe.