Mutex/Once/OnceLock aren't guaranteed to have atomic acquire/release semantics

[briansmith]

Location

https://doc.rust-lang.org/std/sync/struct.Mutex.html, including at least:
https://doc.rust-lang.org/std/sync/struct.Mutex.html#method.lock
https://doc.rust-lang.org/std/sync/struct.Mutex.html#method.try_lock
https://doc.rust-lang.org/std/sync/struct.MutexGuard.html

Summary

See https://marabos.nl/atomics/memory-ordering.html#example-locking, emphasis mine:

Mutexes are the most common use case for release and acquire ordering (see "Locking: Mutexes and RwLocks" in Chapter 1). When locking, they use an atomic operation to check if it was unlocked, using acquire ordering, while also (atomically) changing the state to "locked." When unlocking, they set the state back to "unlocked" using release ordering. This means that there will be a happens-before relationship between unlocking a mutex and subsequently locking it.

Many people are relying on this guarantee, but the documentation for Mutex does not state this guarantee. We should explicitly guarantee this in the documentation for Mutex.

In theory, a mutex could have "consume" semantics instead of acquire/release semantics, such that reads/writes to the "contents" of a mutex (the T in Mutex<T>) and pointers/reference contained therein, are treated as "consume" load/stores (even if Rust's atomics API doesn't support consume ordering). In practice, people have already written much code that relies on the acquire/release semantics.

Or, in theory, a mutex could have SeqCst semantics during lock and unlock, but I think we don't want people to rely on that (though I already have seen code that does, accidentally).

Why does this matter?. Consider:

use std::{
    fs, io,
    os::fd::{IntoRawFd as _, RawFd},
    sync::{
        atomic::{compiler_fence, AtomicI32, Ordering},
        Mutex,
    },
};

fn get_lazy_fd() -> Result<RawFd, io::Error> {
    const FD_UNINIT: i32 = -1; // libstd guarantees a RawFD is never -1.

    // FD isn't stored within the mutex so we can usually use it
    // without locking overhead.
    static FD: AtomicI32 = AtomicI32::new(FD_UNINIT);
    static LOCK: Mutex<()> = Mutex::new(());
    match FD.load(Ordering::Relaxed) {
        fd if fd != FD_UNINIT => Ok(fd), // Common case: Fast, no Acquire at all.
        _ => {
            let _guard = LOCK.lock().unwrap();

            if let Some(fd) = FD.load(Ordering::Relaxed) {
                return fd;
            }

            let fd = fs::File::open("/dev/urandom")?.into_raw_fd();

            FD.store(fd, Ordering::Relaxed);
            
            Ok(fd)
        }
    }
}

We want some official reassurance of the "obvious" fact that the FD.store(fd, Ordering::Relaxed) will be sequenced before the unlocking of the Mutex, i.e. that the mutex is doing an atomic release when unlocking the mutex so that the side effect of storing to FD will be seen by other thread after locking the mutex, which would only be guaranteed if locking the mutex does an atomic aquire that synchronizes with that atomic release.

This would also clarify the more basic fact that in a Mutex<T>, there is nothing special about what's "contained" in the Mutex, as far as synchronization is concerned. (This is unique to rust; C/C++/Posix mutexes don't have the concept of the Mutex containing a value.)

[chorman0773]

I don't think we need to explicitly say "It does an atomic operation".

We can just say

The return from a call to Mutex::lock synchronizes-with a preceeding call to MutexGuard::drop that unlocks that Mutex.

[briansmith]

We can just say

The return from a call to Mutex::lock synchronizes-with a preceeding call to MutexGuard::drop that unlocks that Mutex.

That sounds good. However, do we have a place that says that synchronizes-with implies happens-before? I only see it implicitly in https://doc.rust-lang.org/std/sync/atomic/fn.fence.html:

A fence ‘A’ which has (at least) Release ordering semantics, synchronizes with a fence ‘B’ with (at least) Acquire semantics, if and only if there exist operations X and Y, both operating on some atomic object ‘M’ such that A is sequenced before X, Y is sequenced before B and Y observes the change to M. This provides a happens-before dependence between A and B.

Whereas Ordering::Release and Ordering::Acquire basically indirectly say that AtomicX.store(x, Ordering::Release) and AtomicX::load(Ordering::Acquire) establish a happens-before relationship, but without mentioning synchronized-with or happens-before.

It seems like if we want to use the term synchronizes-with for Mutex, then we should change the definition of Release/Acquire to also use that term, and then somewhere give some kind of definition of synchronizes-with that specifically mentions "happens before."

https://preshing.com/20130823/the-synchronizes-with-relation/#other-ways-to-achieve-synchronizes-with and some other places also mention that spawning a thread synchronizes-with (Ordering::Release) the execution of the closure on the new thread (Ordering::Acquire).

I believe std::sync::Once and OnceLock also need similar language; every call to Once::call_once()/OnceLock::get_andinit().OnceLOck::get_or_try_init() synchronizes-with (Ordering::Acquire) the successful (i.e. non-panicking, non-Err result) completion of the initialization function (Ordering::Release). (Note that the documentation for std::sync::Once seems to say something much stronger right now, which is another issue.)

[veera-sivarajan]

@rustbot label -needs-triage +A-atomic

[chorman0773]

That sounds good. However, do we have a place that says that synchronizes-with implies happens-before?

Yes, the C++ Standard, [intro.races]#9, [intro.races]#10, and [intro.races]#11 https://eel.is/c++draft/intro.races#9.

[alexpyattaev]

Mutex explicitly documented as having what amounts to "side effects" (a.k.a. spooky action at a distance) w.r.t. random atomics does sound somewhat inappropriate for rust in general, especially in environments that might have the machinery to support a mutex, but not necessarily the side-effects that you find on e.g. an ARM implementation of it. Specifically, if you want to enforce a particular ordering w.r.t. a particular atomic, maybe that is what you need to state in the code using an appropriate fence call? Expecting a mutex to implicitly do that for you essentially relies on mutex's internal implementation to work. I might be missing something very important here, of course, but it seems that exposing these internal details into the docs may not be the best idea for future portability of Mutex to newer architectures, unless someone can certify that it is impossible to implement mutex in any other way.

[chorman0773]

I don't think any other implementation exists that is sound, no.
Consider Mutex<Box<T>>. For it to be sound to access the memory inside the Box, every access from all previous attempts at locking the Mutex must happen-before the next time you can access the Box. This implies either synchronizes-with or dependency ordered before at the very least, otherwise you don't get happens-before when you have two different threads.

Being a fence is an overly restrictive specification, but it saying synchronizes-with is required IMO for the API to be sound.

Edit: Yes, Mutex could probably use dependency ordered before, depending on what exactly the preconditions on Send impls are. However, given consume is a hot mess in C++, to the point where it's outright discouraged as of C++17, I don't think we want to use dependency ordering. It may be nice if, in Rust, happens before=simply happens before.

[alexpyattaev]

happens before=simply happens before.

That does indeed sound like "the rust way".

However, from what I can see in c++ specs, some of them claim that memory order consume was in fact re-added in c++20, and its former definition was discouraged since c++17. It is all a bit wonky in there, so I'd just ignore the C++ spec entirely for the purpose of this discussion, especially since llvm has no implementation for consume anyway. What truly matters is what CPUs can or can not do in principle.

With respect to what consume ordering was supposed to do, as far as I can see, it was intended to support specifically use-cases such as rust's Mutex. As a former hardware engineer, I can envision mechanisms that would enable hardware-level simplifications in cases when Mutex-protected area is very small (e.g. one cache line), as I could plausibly do less work then in the case of a full "happens before" synchronization. Arguably, the main benefits would be seen on NUMA machines and GPUs, and we should look for what is done e.g. in CUDA to be certain.

On the other hand, I'd imagine this sort of aggressive optimization, even if we find a way to implement consume "correctly", would break horribly around unsafe code. Compiler would need a far better understanding of what depends on what to not mess this up, and unsafe/interior mutability stuff will not make it easy at all.

As a solution to the doc problem, a generic parameter may be provided to std::Mutex that would toggle whether it uses memory ordering acquire-release or acquire-consume, with acquire-release being the default. This way we are explicit about the side-effects that may occur or relations that can be relied upon, and can document it as such. If/when memory order consume becomes a thing, users can switch to that without breaking the world.