After some recent discussion on the topic of mutex poisoning and notably a blog post that I thought was interesting, thoughtful, exhaustive, yet completely disagreed with, I wanted to comment on the debate myself. To me, poisoning-by-default is one of biggest mistakes that were made in the language, next to some others that are a story for a different time.
The only real argument for not poisoning is “performance” (at the expense of safety). Provide a real-world use-case where you think poisoning is supposedly the bottleneck, and I will (try to) show you that it isn’t.
I like how everyone is pretending writing your own extension trait is not a thing, where you can just add methods like panicking-lock() (with whatever short name you like), but also for backed-off lock (with try_lock() which is one atomic op and time-outs).
I’m having trouble finding any popular alternative implementations of
Mutexthat support poisoning.parking_lot,tokio, andfuturesall have non-poisoning (possibly async) mutexes.In fact, there’s almost no argument for requiring poisoning. There are some fringe cases where poisoning is useful, and in those cases, you could just as easily create a
Poisoning<T>pointer that implements it.As the article says, no other built-in pointer type in Rust support poisoning, yet the same “invalid state” argument can be made for all of them. A
&mut Tcan become temporarily “invalid” (in a sound manner) during a critical block that panics, for example. That pointer could be mutating data that’s eventually dropped during unwinding, does not hold its invariants at this time, and triggers aDropimpl that’s unsound as a result. At the end of the day, the correct way to handle this is to somehow either fix the state during unwinding so that it no longer violates an invariant, or mark it as poisoned, regardless of pointer type.The only real argument for poisoning by default is because it already is the default. It’s a strong argument of course (it would likely be a breaking change to remove it), but it’s not a very solid argument for why it should be a default to begin with.
https://docs.rs/saa/5.6.4/saa/lock/struct.Lock.html
I don’t think performance is the bottleneck. The real reason to not use poisoning mutexes is convenience and developer experience. The standard library should take such considerations heavily into account. A rare use case such as poisoning being requried should not be the default for mutexes in the standard library.
What does that even mean?
The interface concerns are as trivial as it gets:
#![feature(nonpoison_mutex)] #![feature(sync_nonpoison)] use std::sync::Mutex as PMutex; use std::sync::nonpoison::Mutex as NoPMutex; trait LockExt<'a> { type Guard; fn l(&'a self) -> Self::Guard; } impl<'a, T: 'a> LockExt<'a> for PMutex<T> { type Guard = std::sync::MutexGuard<'a, T>; fn l(&'a self) -> Self::Guard { self.lock().expect("non-poisoned") } } impl<'a, T: 'a> LockExt<'a> for NoPMutex<T> { type Guard = std::sync::nonpoison::MutexGuard<'a, T>; fn l(&'a self) -> Self::Guard { self.lock() } } fn main() { let m_p = PMutex::new(false); let m_no_p = NoPMutex::new(false); *m_p.l() = true; *m_no_p.l() = true; dbg!(*m_p.l()); dbg!(*m_no_p.l()); }This is basically is the one thing that shouldn’t consume any mind tokens when it comes to synchronization primitive choice.
I’m not really sure what you’re trying to say here. Are you saying that this is not a problem and everyone should just add this LockExt trait to all their code bases that uses Mutexes? Or…?
Every non-trivial code-base already has a
utilsmodule (sometimes a whole crate even) containing, in part, extension traits (and sometimes macros) to make some usages of some types nicer, in a way that is specific to the codebase (even in this case, a codebase might want a debug or trace log message for example, alongside the panic).It’s a part of the language for a reason. It provides full idiomacity. And this is not limited to (poisoned or otherwise) synchronization primitives.
Talking about a method returning a
Resultlike it’s a big deal must be some sort of splitting-hairs maxxing.And btw, as it just happens, such traits actually already exist, in a crate that is well-known to the ecosystem.
Well I would say this is not at all specific to the codebase - as I said above, it applies to basically all code bases that use mutexes, since poisoning is so rarely relevant.
These two statements are kind of strange to me:
and
If this is such a small deal, then using an external dependency is a terrible solution. That’s like the whole left-pad situation, a dependency to solve an extremely small problem, but something that affects many, many code bases. That’s exactly the use case that the standard library should cover, exactly so that you don’t need to use an external dependency for such a trivial thing!
And yea you could say “duh just write the extension trait in your utils module” but this is just not acknowledging the problem at all and just sidestepping it. That’s not helping anyone.
I was talking about what a codebase would want a non-Result-returning
lock()to look like (the panic message, or forcing abort, extra logging, signal handling…etc).Poisoning (as a negative) is only relevant if we are pretending it matters for performance. And you already conceded that this is not a part of the conversation as far as you’re concerned.
The “rarely relevant” (as in playing any operative role at all) is also pure conjecture, or if we want to be nicer, it’s totally contextual and use-case dependant.
lock_apiprovides a generic interface so you can e.g. seamlessly swipe out primitives fromparking_lotwith primitives fromspin, you know, the whole point behind generic programming!This is in no way comparable with
leftpad. And I wasn’t actually pointing to it as a solution (to the largely non-existing problem). I just wanted to point out that extension traits for synchronisation primitives already exist with wide usage in the ecosystem, since you seemed to not be aware.If you have an aversion to extension traits. Here is a trivial unit struct wrapper solution. You don’t even need to change the
Mutextype name or thelock()method name in your code.use std::ops::{Deref, DerefMut}; use std::sync::Mutex as PMutex; #[repr(transparent)] struct Mutex<T>(PMutex<T>); impl<T> Deref for Mutex<T> { type Target = PMutex<T>; fn deref(&self) -> &Self::Target { &self.0 } } impl<T> DerefMut for Mutex<T> { fn deref_mut(&mut self) -> &mut Self::Target { &mut self.0 } } impl<T> Mutex<T> { fn new(v: T) -> Self { Self(PMutex::new(v)) } fn lock<'a>(&'a self) -> std::sync::MutexGuard<'a, T> { self.0.lock().expect("poisoned") } } fn main() { let m = Mutex::new(false); // oh look, it's a non-Result returning lock() *m.lock() = true; dbg!(*m.lock()); // inner Mutex method API is still available via Deref/DerefMut *m.try_lock().unwrap() = false; dbg!(*m.lock()); }What is the notable relevant problem. Can you state it plainly?
I already told you, convenience, developer experience. Ergonomics basically. None of this is fixed by some code snippet, whether it’s an extension trait or wrapper type. That just makes things worse. This has nothing to do with performance, as I already said.
Can you point the notable relevant inconvenience, bad developer experience, and poor ergonomics developers are suffering from?
Is it all caused by a
lock()method returning aResult?Anything else?