Incorrect unsafe code example in standard library

[TrolledWoods]

The safety clause on the Vec::set_len function says that:

• new_len must be less than or equal to capacity().
• The elements at old_len..new_len must be initialized.

However, the code example for std::ptr::copy violates the second rule

use std::ptr;

unsafe fn from_buf_raw<T>(ptr: *const T, elts: usize) -> Vec<T> {
    let mut dst = Vec::with_capacity(elts);
    dst.set_len(elts);
    ptr::copy(ptr, dst.as_mut_ptr(), elts);
    dst
}

The set_len is called before the values are initialized with the copy.

This could be fixed by either making the safety clause for set_len more inclusive, so that as long as you don't use the vector before initializing the values it's considered safe, or by switching the dst.set_len(elts); and ptr::copy(ptr, dst.as_mut_ptr(), alts); lines around in the example.

[leonardo-m]

Eventually Rust will need pre-conditions and post-conditions to verify statically and automatically that (most) invariants of unsafe functions hold :-)

[TrolledWoods]

Isn't the point of unsafe code that you yourself uphold the invariants?

[lcnr]

yeah, afaict the safety conditions of set_len currently do not guarantee that set_len doesn't look at any elements, so this would be an allowed implementation of set_len afaict.

fn set_len<T>(v: &mut Vec<T>, len: usize) {
    v.len = len;
    if let Some(last) = v.last() {
        let _x: ManuallyDrop<T> = unsafe {
            std::ptr::read(last as *const T as *const ManuallyDrop<T>)
        };
        // put copy of a `T` on the stack, 
        // requires that that value satisfies the language invariants of `T`
    }

We "obviously" do not do this, but I guess cc @rust-lang/wg-unsafe-code-guidelines on what's the right wording here.

[leonardo-m]

Currently yes, but surely that isn't a requirement. Having some mechanical ways to avoid some bugs in unsafe{} sections will be very good, even if it doesn't cover everything. There are people working on the formalization of (unsafe) Rust semantics that have this purposes too.

And in some cases it's not that hard, is well within the capabilities of other research languages, like here:
https://raw.githubusercontent.com/rust-lang/rust/f976b4641b963dbcc1ed4edb155fce5db28807b1/library/alloc/src/vec.rs

The second SAFETY note says:

    // SAFETY: A `Vec`'s pointer is always aligned properly, and
    // the alignment the array needs is the same as the items.
    // We checked earlier that we have sufficient items.
    // The items will not double-drop as the `set_len`
    // tells the `Vec` not to also drop them.

Zig language gives pointers a static alignment annotation in the type system. And probably the annotations of a language like Why3 are enough for the other part.

[TrolledWoods]

@leonardo-m I think this discussion is irrelevant to the issue. What we should be discussing is what to do; change set_len safety guarantees, because "We 'obviously' do not do this", or change the example. I'm leaning towards making the safety guarantees for set_len less strict, because that would allow for more flexible usage of the function.

[TrolledWoods]

However, changing the safety on the set_len would break the invariants of Vec, and if you want to do crazy things with a vector you can use into_raw_parts and from_raw_parts. In my experience rust doesn't like invariants to be broken even for just short periods of time inside unsafe code, so just fixing the example may be a better choice that fits better with how rust works.

[Lokathor]

The real heart of things is that code within the module is allowed to break the safety invariant as long as they don't break the validity invariant. As long as things are all put back together properly by the end of the function.

[TrolledWoods]

A code example is not within the module

[Lokathor]

Ah my mistake! I thought this was a vector construction method on the vec type.

If this is anywhere else at all then the two lines need to be swapped so that set_len comes after the initialization.

[RalfJung]

What we should be discussing is what to do; change set_len safety guarantees, because "We 'obviously' do not do this", or change the example. I'm leaning towards making the safety guarantees for set_len less strict, because that would allow for more flexible usage of the function.

Exactly.

Give that Vec already makes plenty of guarantees to unsafe code, I think it would be reasonable to say that set_len itself is safe no matter which length you set, but before you use the vector again (i.e., before you call any vector operation, including drop), you must ensure that the given invariants hold.

However, to make the example code correct, we'd additionally have to say that as_mut_ptr is always safe to call, despite the "do not call any other operation" in set_len.

All of this is a roundabout way to say that set_len can violate the safety invariant (or "library invariant") of the type, and that as_mut_ptr is okay to call even when the safety invariant is currently violated.

[lcnr]

I think it would be reasonable to say that set_len itself is safe no matter which length you set, but before you use the vector again (i.e., before you call any vector operation, including drop), you must ensure that the given invariants hold.

Note that we currently have a debug assert in set_len for len <= capacity, so setting the len to something greater than the capacity can currently panic which probably can cause unsoundness as it would be unexpected.

If we want to change the safety guarantees here we probably should remove that debug_assert (even if they aren't actually part of the std used with rustup afaik)

[RalfJung]

@lcnr good point. Or we could say that a length larger than the capacity is insta-UB (library UB).