22 KiB

Raw Permalink Blame History

Restrictions

Start Date: 2022-10-09
RFC PR: rust-lang/rfcs#3323
Rust Issue: rust-lang/rust#105077

Summary

You can write pub impl(crate) trait Foo {}, which limits the ability to implement the trait to the crate it is defined in. Similarly, you can write pub struct Foo(pub mut(crate) u8); and pub struct Foo { pub mut(crate) foo: u8 }, which limits the ability to mutate the u8 to crate. Outside of the declared scope, implementing the trait or mutating the field is not allowed. If no restriction is specified, the ability to implement or mutate is uninhibited.

Motivation

Currently, a trait being visible (and nameable) in a given location implies that you are able to implement it. However, this does not mean that you want anyone to implement it. It is reasonable to want a trait to only be implemented by certain types for a variety of reasons. This is commonly referred to as a "sealed trait", and is frequently simulated by using a public trait in a private or restricted module.

Similarly, a field being visible currently implies that you are able to mutate it. Just as with traits being able to be implemented anywhere, this is not always what is wanted. The semantic correctness of a field may depend on the value of other fields, for example. This means that making fields public, while acceptable for read access, is not acceptable for write access. Limiting the ability to mutate a field to a certain scope is desirable in these situations, while still allowing read access everywhere else.

Guide-level explanation

Restrictions limit what you are allowed to do with a type. In this sense, visibility is a restriction! The compiler stops you from using a private type, after all. #[non_exhaustive] is also a restriction, as it requires you to have a wildcard arm in a match expression. Both of these are used on a daily basis by countless Rust programmers.

Restrictions are a powerful tool because the compiler stops you from doing something you are not allowed to do. If you violate a restriction by using unsafe trickery, such as transmuting a type, the resulting code is unsound.

So why do we need restrictions? In fact, they are incredibly important. Those that have been around a while will remember a time before #[non_exhaustive]. Standard practice at that point in time was to include a #[doc(hidden)] __NonExhaustive variant on enums and a private non_exhaustive: () field on structs. There are two problems with this approach. First, the variant or field exists! Yes, that is obvious, but it is worth noting that the user can still match exhaustively. Second, the dummy variant has to be handled even within the crate that defined it. With the #[non_exhaustive] restriction, this is not the case.

`impl`-restricted traits

It is very common for a library to want to have a trait that exists but only have it be implemented for the types they want. It is so common, in fact, that there are official guidelines on how to do this! The pattern is typically referred to as a "sealed trait". Here is a modified example from the guidelines:

/// This trait is sealed and cannot be implemented for types outside this crate.
pub trait Foo: private::Sealed {
    // Methods that the user is allowed to call.
    fn bar();
}

// Implement for some types.
impl Foo for usize {
    fn bar() {}
}

mod private {
    pub trait Sealed {}

    // Implement for those same types, but no others.
    impl Sealed for usize {}
}

That is a fair amount of code to say "you cannot implement Foo"! This works because it is permitted to have a public item (Sealed) in a private module (private). More specifically, Sealed is public, but users in another crate are unable to name the trait. This effectively makes the trait private, assuming it is not used in other manners. It would be far nicer if you could just write:

pub impl(crate) trait Foo {
    fn bar();
}

impl Foo for usize {
    fn bar() {}
}

Note that there is neither a Sealed trait nor a private module here. The ability to implement Foo is restricted by the compiler. It knows this because we used impl(crate) — the new syntax introduced here. Just as pub accepts a module path, impl does the same. This means that impl(super) and impl(in path::to::module) are also valid. Using the impl keyword in this position is a natural extension of the existing visibility syntax. The example above would restrict the ability to implement the trait to the defining crate. If we used impl(super) instead, it would be restricted to the parent module. If we used impl(in path::to::module), it would be restricted to the specified module. Any attempt to implement the trait outside of these modules will error. For example, this code:

pub mod foo {
    pub mod bar {
        pub(crate) impl(super) trait Foo {}
    }

    // Okay to implement `Foo` here.
    impl bar::Foo for i8 {}
}

impl foo::bar::Foo for u8 {} // Uh oh! We cannot implement `Foo` here.

could result in the following error:

error: trait cannot be implemented outside `foo`
  --> $DIR/impl-restriction.rs:13:1
   |
LL |         pub(crate) impl(super) trait Foo {}
   |                    ----------- trait restricted here
...
LL | impl foo::bar::Foo for u8 {}
   | ^^^^^^^^^^^^^^^^^^^^^^^^^

error: aborting due to previous error

There are benefits to having this restriction built into the language. First, it expresses the intent of the author more clearly. Documentation can automatically show that the implementation is restricted, and the compiler can emit better diagnostics when someone tries to implement Foo. Another benefit is that it is no longer possible to accidentally implement Sealed for a type but not Foo. This is a very easy mistake to make, and it is difficult to notice. With the new syntax, you will only have one trait to worry about.

`mut`-restricted fields

Have you ever wanted to have read-only fields in Rust? C++, C#, Java, TypeScript, Kotlin, and Swift all have them in some form or another! In Rust, it is feasible to go one step further and have fields that are only mutable within a certain module. Said another way, you can mutate it but other people cannot. This is useful for a number of reasons. For example, you may have a struct whose values are always semantically in a given range. This occurs in time:

pub struct Time {
    hour: u8,
    minute: u8,
    second: u8,
    nanosecond: u32,
}

The author of time would love to have these fields public. However, they do not want users to be able to change the values, as that would violate the invariants of the type. As a result they currently have to keep the fields private and write "getter" methods. What if, instead, they could add mut(crate) to a field, just like pub(crate)? This would allow them to write:

pub struct Time {
    pub mut(crate) hour: u8,
    pub mut(crate) minute: u8,
    pub mut(crate) second: u8,
    pub mut(crate) nanosecond: u32,
}

This would mean that the fields are mutable within time, but not outside. This avoids the need to write getters for fields that already exist. While for a type like Time this is not a big deal, having access to fields directly instead of through getters can help with borrow checking. This is because the compiler is smart enough to know that field accesses cannot overlap, but it does not know this solely from the function signature of getters.

While there is the readonly crate, this approach has its drawbacks. Namely, the type cannot implement Deref: it already does because of this macro. It is not possible to have only some fields be read-only: Deref is all-or-nothing. It is not possible to make the fields mutable only within a certain module: Deref is a trait that cannot be implemented only in certain locations. Furthermore, readonly does not in any way help with borrow checking. While useful in some situations, it is by no means a complete solution.

Where does a mutation occur?

There is one major question: what even counts as a mutation? This is not as straightforward as you might think. If you write

let mut x = 5;
let y = &mut x;
*y = 6;

It is without question that a mutation occurs. But where? Does it occur on the second or third line? In this example, it would not matter, but it is easy to imagine passing a mutable reference to a function that then mutates the value. There, it is not clear where the mutation occurs. The answer is that the mutation occurs on the line where the reference is taken. This is the choice that makes the most sense from the perspective of the user.

fn foo<T>(x: &mut T, value: T) {
    if random() {
        *x = value;
    }
}

let mut x = 5;
foo(&mut x, 6);

Here, x is mutably borrowed on the final line, but the value is changed in memory inside the if block. You might say, logically, that the mutation occurs inside the if block. But if we use this definition, then we could not know about the mutation until after monomorphization. Errors generated post-monomorphization are generally frowned upon, as it happens quite late in the compilation process. But consider this: what if x is not actually mutated within the body of foo? Now we have a window into what actually happens inside the function, and it is something that is not stated in the function signature. Not great. In this specific example, it is not even deterministic! Because of this, it is quite literally impossible to know whether x is actually mutated inside a given function. As a result we have no choice: the error must be generated at the point where the reference is taken.

Okay, we solved that problem. We know that the mutable use happens on the final line. But what about this?

let x = Cell::new(5);
x.set(6);

Rust has interior mutability, which is what we are using here. x is not declared mutable, and it does not need to be. This is the purpose of interior mutability, by definition. But it introduces a key question: where is the mutation? The answer is that it is not a mutation for the purposes of this restriction. This is not because the value is not changed: it is. Rather, it is the logical result of the semantics of mut restrictions and where errors must occur (as described after the previous example). If errors are emitted at the point where the mutable reference is created, then there can be no such error here, as no mutable reference is ever created. Cell::set is a method that takes &self, not &mut self. Interior mutability is not special-cased; the only way to work around this would be to make even non-mutable reference to a type with interior mutability considered a mutation. Consequently, you could never have a reference to a type containing a mut-restricted, interior-mutable field. This is unacceptable, so interior mutability cannot be considered a mutation for the purposes of this restriction. Interfaces that wish to restrict even interior mutability of a field should avoid exposing it as a public field with private mutability.

`struct` expressions are not allowed

Given that the most common use for for mut-restricted fields is to ensure an invariant, it is important that the invariant be enforced. Consider the previous definition of Time. If you could write

Time {
    hour: 32,
    minute: 0,
    second: 0,
    nanosecond: 0,
}

then the invariant would be violated, as there are only 24 hours in a day (numbered 0–23). Given that the invariant is not enforced by the type system, it cannot be enforced at all in this case. As a result, we have no choice but to disallow struct expressions for types with mut-restricted fields, in scopes where any fields are mut-restricted. This applies even when functional update syntax is used, as invariants can rely on the value of other fields.

Note that despite the name, struct expressions are not limited to structs. They are used to initialize enum variants and unions as well. For enums and unions, this restriction only applies to the specific variant being constructed. For example, the following is allowed:

pub enum Foo {
    Alpha { mut(crate) x: u8 },
    Beta { y: u8 },
}

// In another crate:
Foo::Beta { y: 5 };

In this example, Foo::Alpha { x: 5 } is allowed when it is in the same crate as Foo. This is because x is not restricted within this scope, so the field can be freely mutated. Because of this, the previous concern about upholding invariants is not applicable.

Reference-level explanation

Syntax

Using the syntax from the reference for structs, the change needed to support mut restrictions is quite small.

StructField :
   OuterAttribute*
   Visibility?
+  MutRestriction?
   IDENTIFIER : Type

TupleField :
   OuterAttribute*
   Visibility?
+  MutRestriction?
   Type

+MutRestriction :
+   mut ( crate )
+   | mut ( self )
+   | mut ( super )
+   | mut ( in SimplePath )

Trait definitions need a similar change to the syntax for traits to accommodate impl restrictions.

Trait :
   unsafe?
+  ImplRestriction?
   trait IDENTIFIER
   GenericParams? ( : TypeParamBounds? )? WhereClause? {
     InnerAttribute*
     AssociatedItem*
   }

+ImplRestriction :
+   impl ( crate )
+   | impl ( self )
+   | impl ( super )
+   | impl ( in SimplePath )

Essentially, mut and impl have the same syntax as pub, just with a different keyword. Using the keyword without providing a path is not allowed.

Behavior

The current behavior of pub is that pub makes something visible within the declared scope. If no scope is declared (such that it is just pub), then the item is visible everywhere. This behavior is preserved for impl and mut. When a restriction is used, the behavior is allowed only within the declared scope. While in most cases the default visibility is private, pub is default in some cases, namely enum variants, enum fields, and trait items. impl and mut will have a consistent default: when omitted entirely, the scope is inherited from pub. This is both what is most convenient and is what is required for backwards compatibility with existing code.

When an ImplRestriction is present, implementations of the associated trait are only permitted within the designated path. Any implementation of the trait outside this scope is a compile error. When a MutRestriction is present, mutable uses of the associated field are only permitted within the designated path. Any mutable use of the field outside the scope is a compile error. Further, a struct, union, or enum variant containing fields with an associated MutRestriction may not be constructed with struct expressions unless all fields are unrestricted in the present scope. This is the case even if the field is not directly declared, such as when functional record updates are used.

"Mutable use" in the compiler

The concept of a "mutable use" already exists within the compiler. This catches all situations that are relevant here, including ptr::addr_of_mut!, &mut, and direct assignment to a field, while excluding interior mutability. As such, formal semantics of what constitutes a "mutable use" are not stated here.

Interaction with `trait` aliases

Trait aliases cannot be implemented. As such, there is no concern about compatibility between the impl restriction and trait aliases.

Drawbacks

Additional syntax for macros to handle
More syntax to learn
While unambiguous to parse, trait impl(crate) Foo could be confusing due to its similarity to impl Foo.

Alternatives

impl and mut restrictions could be attributes, similar to #[non_exhaustive].
- The proposed syntax could by syntactic sugar for these attributes.
Visibility could be altered to accept restrictions as a type of parameter, such as pub(crate, mut = self). This is not ideal because restrictions are not permitted everywhere visibility is. As a result, any errors would have to occur later in the compilation process than they would be with the proposed syntax. It would also mean macro authors would be unable to accept only syntax that would be valid in a given context. Further, some positions such as enum variants do not semantically accept a visibility, while they do accept a restriction.
The current syntax separates the mut/impl keyword from the scope of the restriction. This produces verbose syntax. Many users may want similar restrictions. Could we provide a simpler syntax if we provided less flexibility? Would a new keyword or two help? We could choose a syntax with less flexibility and verbosity but more simplicity. For instance, sealed or readonly.

Prior art

The readonly crate simulates immutable fields outside of the defining module. Types with this attribute cannot define Deref, which can be limiting. Additionally, it applies to all fields and within the defining crate. The advantages of native read-only fields relating to borrow checking also do not apply when using this crate.
The derive-getters and getset crates are derive macros that are used to generate getter methods. The latter also has the ability to derive setters. This demonstrates the usefulness of reduced syntax for common behavior. Further, getset allows explicitly setting the visibility of the derived methods. In this manner, it is very similar to the ability to provide a path to the mut restriction.
The ability to restrict implementations of a trait can be simulated by a public trait in a private module. This has the disadvantage that the trait is no longer nameable by external users, preventing its use as a generic bound. Current diagnostics, while technically correct, are unhelpful to downstream users.
Various other languages have read-only fields, including C++, C#, Java, TypeScript, Kotlin, and Swift.
Users of many languages, including Rust, regularly implement read-only fields by providing a getter method without a setter method, demonstrating a need for this.

Unresolved questions

Should an "unnecessary restriction" lint be introduced? It would fire when the restriction is as strict or less strict than the visibility. This warning could also be used for pub(self).
- Does this necessarily have to be decided as part of this RFC?
How will restrictions work with macro_rules! matchers? There is currently a vis matcher, but it is likely unwise to add a new matcher for each restriction.
- The proposed syntax cannot be added to the vis matcher, as it does not current restrict the tokens that can follow. For this reason, it could break existing code, such as the following example.
```
macro_rules! foo {
    ($v:vis impl(crate) trait Foo) => {}
}

foo!(pub impl(crate) trait Foo);
```
- A restriction matcher could work, but restrictions are not the same everywhere.
- mut_restriction and impl_restriction are relatively long.
What is the interaction between stability and restrictions?
- Suggestion: Visibility is an inherent part of the item; restrictions should be as well. Metadata can be added in the future indicating when an item had its restriction lifted, if applicable. The design for this is left to the language team as necessary. A decision does not need to be made prior to stabilization, as stability attributes are not stable in their own right.
Should the in syntax be permitted for restrictions? Including it is consistent with the existing syntax for visibility. Further, the lack of inclusion would lead to continued use of the workaround for impl. For mut, there is no workaround. The syntax is not used often for visibility, but it is very useful when it is used.
Should struct expressions be disallowed?
- Where would it be desirable to prohibit mutability after construction, but still permit construction with unchecked values?
Should a simpler syntax be provided for common cases? For instance, sealed or readonly. A different syntax altogether could be used as well.

Future possibilities

Explicitly sealed/exhaustive traits could happen in the future. This has the ability to impact coherence, such that other crates could rely on the fact that the list of implementations is exhaustive. As traits would default to unsealed, this does not have be decided now.
Trait items could gain proper visibility and/or restrictions of their own. This would allow private and/or defaulted trait items that cannot be overridden.
Set-once fields could potentially occur in the future. Functionally, this would be "true" read-only fields, in that they can be constructed but never mutated. They are not included in this proposal as the use case is nor clear, nor is there an immediately obvious syntax to support this.
The default could be changed in a future edition, such as to make pub field: Type be only mutable within the module rather than mutable everywhere. This seems unlikely, as it would be an incredibly disruptive change, and the benefits would have to be significant.
Syntax such as impl(mod) could be added for clarity as an alternative to impl(self).
impl and mut could be usable without a path if deemed necessary. This behavior would be identical to omitting the keyword entirely.
mut could be placed on the struct or variant itself, which would be equivalent to having the same restriction on each field. This would avoid repetition.
Trait implementations could be restricted to being used within a certain scope.

22 KiB Raw Permalink Blame History Unescape Escape