[WIP 0/3] Memory model and atomic API in Rust

[comex]

On Mar 25, 2024, at 8:49 PM, Linus Torvalds <torvalds at linux-foundation.org> wrote:

> But you should _start_ the design of your language memory model around
> the unsafe "raw atomic access operations" model.
> 
> Then you can use those strictly more powerful operations, and you
> create an object model *around* it.

To some extent Rust does this already, unlike C++.

C++ allows atomics to be implemented using locks.  Partly for this reason,
`std::atomic<T>` is documented as not necessarily having the same
representation as `T` [1].  C++ also has strict aliasing, so even if those types
do have the same representation, you still can't cast `T *` to
`std::atomic<T> *`.

But Rust atomics are lower-level.  First, they are guaranteed lock-free [2].
Second, they are documented as having "the same in-memory representation as the
underlying" type [3].  (They also usually have the same alignment, except on
x86 where u64 is only 4-byte aligned but AtomicU64 of course needs to be 8-byte
aligned.)  Meanwhile, Rust intentionally lacks strict aliasing.

Combined, this means it's perfectly legal in Rust to cast e.g. `&mut u32` to
`&AtomicU32` and perform atomic accesses on it.  Or the same with u64/AtomicU64
if you know the pointer is validly aligned.  This is by design; the Atomic
types' methods are considered the official way to perform atomic operations on
arbitrary memory, making it unnecessary to also stabilize 'lower-level'
intrinsics.

That said, there *are* currently some holes in Rust's atomics model, based on
the fact that it's mostly inherited from C++.  From the documentation:

> Since C++ does not support mixing atomic and non-atomic accesses, or
> non-synchronized different-sized accesses to the same data, Rust does not
> support those operations either. Note that both of those restrictions only
> apply if the accesses are non-synchronized.
https://doc.rust-lang.org/std/sync/atomic/index.html

There are some open issues around this:

- "How can we allow read-read races between atomic and non-atomic accesses?"
  https://github.com/rust-lang/unsafe-code-guidelines/issues/483

  > [..] I do think we should allow such code. However, then we have to change
  > the way we document our atomics [..]

- "What about: mixed-size atomic accesses"
  https://github.com/rust-lang/unsafe-code-guidelines/issues/345"

  > Apparently the x86 manual says you "should" not do this [..] It is unclear
  > what "should" means (or what anything else here really means, operationally
  > speaking...)

[1] https://eel.is/c++draft/atomics#types.generic.general-3
[2] https://doc.rust-lang.org/std/sync/atomic/index.html#portability
[3] https://doc.rust-lang.org/nightly/std/sync/atomic/struct.AtomicU64.html