Public API and Invariants

Reinventing From Scratch — Box<T>

Chapter 3 — Public API Design & Invariants

3.1 What we’ll support (sized T)

  • Constructors: new
  • Raw access & ownership transfer: as_ptr, as_mut_ptr, into_raw, from_raw, leak
  • Ergonomics: Deref, DerefMut
  • Destruction: Drop

3.2 Invariants (tattoo these)

  • ptr is either:
    • valid, aligned pointer to an initialized T, or
    • null only after into_raw disarms the destructor.
  • On drop: drop_in_place → dealloc (in this order).
  • from_raw is only fed pointers produced by into_raw (same type, allocator).

3.3 Skeleton type

pub struct MyBox<T> {
    ptr: *mut T,
}

3.4 Safety boundaries

All unsafety is hidden behind methods that establish or consume invariants. Public methods are safe to call iff the documented contracts are respected.

📝 Doc-first design: For every unsafe method, write its contract in English before writing code.

Exercises

  1. Write a one-paragraph safety contract for from_raw.
  2. Which public methods of MyBox<T> should be unsafe, if any?

Deep Dive: Ownership Proofs, Drop Order, and DST Considerations

A. Formal Invariants for MyBox<T> (Sized)

  • B1 (Pointer Validity): ptr is either null only after into_raw or a valid, properly aligned pointer to initialized T.
  • B2 (Single Drop): The destructor of T is invoked exactly once if and only if ptr is non-null at Drop time.
  • B3 (Dealloc after Drop): dealloc(layout_of::<T>()) is called exactly once, and only after drop_in_place.
  • B4 (From/Into Raw Consistency): from_raw only accepts pointers produced by into_raw of the same type/allocator; mixing allocators is UB.
  • B5 (No References to Uninit): No &/&mut references are created before ptr::write initializes the allocation.

B. Proof Sketches

B.1 Single Drop — Drop checks for null and calls drop_in_place once; into_raw nulls out ptr and forgets self, preventing Drop from running on a live value.
B.2 No Use-After-Free — Deallocation happens only after the destructor; references returned by Deref are derived from a live ptr and never stored beyond the box’s lifetime.
B.3 Panic Safety — If constructor panics before publishing, no ownership is established; if Drop panics (should not), process aborts, avoiding double-unwind corruption.

C. DST Box Notes

  • Slices (Box<[T]>): store length; the fat pointer (data, len) enables correct deallocation.
  • Box<str>: same as [u8] with UTF‑8 invariant; length in metadata.
  • Box<dyn Trait>: fat pointer (data, vtable); the vtable encodes drop and size/alignment; std uses compiler magic for correct layout.

D. Interop Patterns

  • FFI Ownership Transfer: into_raw -> C takes ownership; C must call back into Rust with from_raw or a custom free.
  • Leaking Globals: leak returns 'static reference, acceptable for process lifetime singletons; document intent.

E. Debugging

  • Double Drop: look for *p assignment instead of ptr::write on uninitialized memory.
  • Mismatched Layout: using dealloc with wrong Layout causes heap corruption; keep Layout::new::<T>() paired.

F. Exercises

  1. Implement try_new returning Result<MyBox<T>, AllocError>.
  2. Add into_inner(self) -> T by ptr::read and skipping dealloc? Explain why you must still dealloc after moving T.
  3. Implement MyVec::into_boxed_slice that hands RawVec buffer to a Box<[T]> safely.

FAQ (Extended)

Q: Does Box<T> guarantee a stable address? A: Yes, the pointee’s address is stable for the life of the box; moving the box moves only the handle.
Q: Why ptr::write not *p = value? A: The latter reads/drops the previous contents (uninitialized), which is UB.
Q: Can Box<T> be null? A: By design, standard Box<T> is non-null; our MyBox may set ptr = null only as a consumed sentinel post-into_raw.
Q: Is Pin<Box<T>> needed for stable address? A: Not for stability; Pin is for forbidding moves via the API.