Fix BTreeMap UB #56648
Fix BTreeMap UB #56648
Conversation
|
r? @sfackler (rust_highfive has picked a reviewer for you, use r? to override) |
rust-highfive
added
the
S-waiting-on-review
|
Would it be worth putting NodeHeader fields in the node types directly to avoid accidentally desynchronizing them? The K2 shenanigans seem unfortunate but I'll defer to your expertise that they're necessary :) |
Unfortunately that will increase their size, because there'll be padding after |
|
Ah right. |
Co-Authored-By: RalfJung <[email protected]>
I'm always open for suggestions for better solutions ;) |
|
@bors r+ |
|
📌 Commit d9c64e5 has been approved by |
bors
added
S-waiting-on-bors
Fix BTreeMap UB BTreeMap currently causes UB by created a shared reference to a too-small allocation. This PR fixes that by introducing a `NodeHeader` type and using that until we really need access to the key/value arrays. Avoiding run-time checks in `into_key_slice` was somewhat tricky, see the comments embedded in the code. I also adjusted `as_leaf_mut` to return a raw pointer, because creating a mutable reference asserts that there are no aliases to the pointee, but that's not always correct: We use `as_leaf_mut` twice to create two mutable slices for keys and values; the second call overlaps with the first slice and hence is not a unique pointer. Fixes rust-lang#54957 Cc @nikomatsakis @gankro
Rollup of 20 pull requests Successful merges: - #53506 (Documentation for impl From for AtomicBool and other Atomic types) - #56343 (Remove not used mod) - #56439 (Clearer error message for dead assign) - #56640 (Add FreeBSD unsigned char platforms to std::os::raw) - #56648 (Fix BTreeMap UB) - #56672 (Document time of back operations of a Linked List) - #56706 (Make `const unsafe fn` bodies `unsafe`) - #56742 (infer: remove Box from a returned Iterator) - #56761 (Suggest using `.display()` when trying to print a `Path`) - #56781 (Update LLVM submodule) - #56789 (rustc: Add an unstable `simd_select_bitmask` intrinsic) - #56790 (Make RValue::Discriminant a normal Shallow read) - #56793 (rustdoc: look for comments when scraping attributes/crates from doctests) - #56826 (rustc: Add the `cmpxchg16b` target feature on x86/x86_64) - #56832 (std: Use `rustc_demangle` from crates.io) - #56844 (Improve CSS rule) - #56850 (Fixed issue with using `Self` ctor in typedefs) - #56855 (Remove u8 cttz hack) - #56857 (Fix a small mistake regarding NaNs in a deprecation message) - #56858 (Fix doc of `std::fs::canonicalize`) Failed merges: - #56741 (treat ref-to-raw cast like a reborrow: do a special kind of retag) r? @ghost
| // ever take a pointer past the first key. | ||
| static EMPTY_ROOT_NODE: LeafNode<(), ()> = LeafNode { | ||
| static EMPTY_ROOT_NODE: NodeHeader<(), ()> = NodeHeader { | ||
| parent: ptr::null(), | ||
| parent_idx: MaybeUninit::uninitialized(), |
There was a problem hiding this comment.
Uninitialized in a static is surprising, isn't it, it doesn't have much purpose other than if miri can potentially use it in analysis? Can't we easily just initialize this at no cost, since it's in a static?
There was a problem hiding this comment.
The header is also used at run-time, and this data is not always initialized, so it needs to have type MaybeUninit.
I don't see the point of trying to initialize this to anything else.
There was a problem hiding this comment.
The only point would be to code defensively, sort of like the safe practices one learns to use in C. Requires saying that a bug with use of uninit value would be worse than a bug where we access an initialized but bogus value. (So this is where Miri comes in? It could diagnose the former better!?)
This is absolutely not a big deal, but I like to remember that we need to be defensive in Rust and not just have one line of defence.
There was a problem hiding this comment.
If MaybeUninit::zeroed() was const fn I'd be happy to r+ a PR that uses that, but currently I don't see a good way to do anything more defensive. Do you have a concrete suggestion?
There was a problem hiding this comment.
if Maybeuninit::new(0) would work, that would be it, but that's all. I think it works?
There was a problem hiding this comment.
Do you want to try if it works?
There was a problem hiding this comment.
I have only one project in rust-lang/rust right now, that's #56440, so I can try after that.
BTreeMap currently causes UB by created a shared reference to a too-small allocation. This PR fixes that by introducing a
NodeHeadertype and using that until we really need access to the key/value arrays. Avoiding run-time checks ininto_key_slicewas somewhat tricky, see the comments embedded in the code.I also adjusted
as_leaf_mutto return a raw pointer, because creating a mutable reference asserts that there are no aliases to the pointee, but that's not always correct: We useas_leaf_muttwice to create two mutable slices for keys and values; the second call overlaps with the first slice and hence is not a unique pointer.Fixes #54957
Cc @nikomatsakis @gankro