[dwattttt]

Relative "number of unsafe keywords" or "lines inside unsafe blocks" isn't a good metric.

It's unsafe to call a C library that gives you a raw pointer, that can be a single line. It's unsafe to use that pointer, that could be a second single line. Carrying that pointer around, the data structures it's in, that's all safe, and doesn't implicate lifetime checking at all, so Rust will let you do silly things with the actual lifetime.

A better metric would be absolute unsafe keywords (because each carries a need to review), or "code in a module that uses unsafe anywhere" vs total code, because module boundaries isolate unsafe.

[NohatCoder]

Yeah, one needs to understand that "unsafe" does not mark which parts of the code are actually unsafe, it simply marks parts where the compiler ignores parts of its rule set. But the implications can crop up anywhere, there is no guarantee that a resulting use-after-free or similar can only happen inside the unsafe blocks.

In short, if you use an unsafe block, then potentially any part of your code is unsafe.

[josephg]

> In short, if you use an unsafe block, then potentially any part of your code is unsafe.

The way I think about it is that the unsafe keyword is a promise that you’re going to maintain the safety invariants yourself, manually. The program is memory correct if it correctly maintains a certain set of invariants. Unsafe punts responsibility over those invariants to the programmer. If you use unsafe and mess up, the program may be in an invalid state. Yes - it might crash, anywhere. But the bug is still almost always in an unsafe block.

It’s often possible to make fully safe wrappers around unsafe code which maintains all those invariants manually. (Either statically or dynamically.). A lot of the rust standard library does this. For example, Vec, Box and slice all use unsafe code internally to create safe APIs.

[DSMan195276]

> But the bug is still almost always in an unsafe block.

That's entirely dependent on how you write your Rust code. If you're derefing an invalid pointer then the bug is usually in how you calculated that pointer value, but the only part that actually requires 'unsafe' is the deref, not the bugged pointer calculation.

Now in properly written Rust code you should be marking all of that code as 'unsafe' in that case and documenting what invariants need to be maintained, but that's entirely on you to do. The only part the compiler actually enforces is that you mark the specific spots where you make use of the operations that 'unsafe' allows.

[josephg]

> The only part the compiler actually enforces is that you mark the specific spots where you make use of the operations that 'unsafe' allows.

Yes, I wish there was a way to mark code as unsafe without also allowing unsafe operations. Its quite common I have some "safe" code that generates values which are eventually used in an unsafe block. If the "safe" code is wrong, my unsafe code will fail in memory-unsafe ways. But there's currently no way to annotate this sort of "safe" code. Rust provides the unsafe keyword - but that keyword is generally reserved for code which needs to actually make unsafe operations (like derefing pointers or calling other unsafe functions).

[DSMan195276]

Yeah I agree with that as well, the unsafe keyword takes on a double-duty which makes it unclear how it's actually intended to be used. The idea that you might mark a function that's only made of safe code as 'unsafe' is not that obvious and not always agreed upon, I've seen plenty of debating about it.

The fact that 'unsafe' also enables all unsafe operations in the function body just makes it even messier, since you definitely _don't_ want that unless the entire body is really nothing but unsafe operations. Thus the "mark a safe function as unsafe" has a clear downside since it allows all the unsafe operations you didn't want to use.

[dwattttt]

> The fact that 'unsafe' also enables all unsafe operations in the function body just makes it even messier

This changed in the 2024 edition to now fire a warning lint, asking you to wrap your unsafe operations in an unsafe block, even inside an unsafe function.

[dwattttt]

I think this point gets missed too often in the face of technical things the Rust compiler does.

There was a world where everyone just used the unsafe keyword everywhere, and Rust code crashed & had memory issues all the time.

The big thing Rust did wasn't invent borrow checking or memory safety; it's the ease of use, the defaults, and the social aspect of "if unsafe is used, a memory bug is in the unsafe part".

[dwattttt]

> it simply marks parts where the compiler ignores parts of its rule set

This is a common misconception, in a literal interpretation. It doesn't invalidate your point, but since people get the wrong impression: unsafe doesn't turn off any of the checks the Rust compiler does.

It allows you to perform 5 additional operations, and that's it. Using those operations wrong is what breaks the safety promises of the language. As an example, you could dereference a raw pointer and tell the Rust compiler it lives forever, when it's really a pointer to an object that's about to be freed.

[randysalami]

Right, so if I understand correctly, to make the code fully safe, it can be impossible? Because you have underlying dependencies which are unsafe. The best you can do is make the handling of the unsafe as safe and as restricted as possible. But I can imagine some of these are way more difficult than others, and in programming the last 10% can be the last 90% of the work. Alas in terms of perceptions, the argument works well.

[dwattttt]

It's not as pessimistic as it sounds. The task of "make the handling of the unsafe as safe as possible" can be quite easy; in some cases, hiding the unsafe in a module and only exposing a safe wrapper means there's only a small amount of code you have to review.