Skip to content

Latest commit

 

History

History
1164 lines (940 loc) · 41.5 KB

0769-sound-generic-drop.md

File metadata and controls

1164 lines (940 loc) · 41.5 KB

History

2015.09.18 -- This RFC was partially superceded by RFC 1238, which removed the parametricity-based reasoning in favor of an attribute.

Summary

Remove #[unsafe_destructor] from the Rust language. Make it safe for developers to implement Drop on type- and lifetime-parameterized structs and enum (i.e. "Generic Drop") by imposing new rules on code where such types occur, to ensure that the drop implementation cannot possibly read or write data via a reference of type &'a Data where 'a could have possibly expired before the drop code runs.

Note: This RFC is describing a feature that has been long in the making; in particular it was previously sketched in Rust Issue #8861 "New Destructor Semantics" (the source of the tongue-in-cheek "Start Date" given above), and has a prototype implementation that is being prepared to land. The purpose of this RFC is two-fold:

  1. standalone documentation of the (admittedly conservative) rules imposed by the new destructor semantics, and

  2. elicit community feedback on the rules, both in the form they will take for 1.0 (which is relatively constrained) and the form they might take in the future (which allows for hypothetical language extensions).

Motivation

Part of Rust's design is rich use of Resource Acquisition Is Initialization (RAII) patterns, which requires destructors: code attached to certain types that runs only when a value of the type goes out of scope or is otherwise deallocated. In Rust, the Drop trait is used for this purpose.

Currently (as of Rust 1.0 alpha), a developer cannot implement Drop on a type- or lifetime-parametric type (e.g. struct Sneetch<'a> or enum Zax<T>) without attaching the #[unsafe_destructor] attribute to it. The reason this attribute is required is that the current implementation allows for such destructors to inject unsoundness accidentally (e.g. reads from or writes to deallocated memory, accessing data when its representation invariants are no longer valid).

Furthermore, while some destructors can be implemented with no danger of unsoundness, regardless of T (assuming that any Drop implementation attached to T is itself sound), as soon as one wants to interact with borrowed data within the fn drop code (e.g. access a field &'a StarOffMachine from a value of type Sneetch<'a> ), there is currently no way to enforce a rule that 'a strictly outlive the value itself. This is a huge gap in the language as it stands: as soon as a developer attaches #[unsafe_destructor] to such a type, it is imposing a subtle and unchecked restriction on clients of that type that they will not ever allow the borrowed data to expire first.

Lifetime parameterization: the Sneetch example

If today Sylvester writes:

// opt-in to the unsoundness!
#![feature(unsafe_destructor)]

pub mod mcbean {
    use std::cell::Cell;

    pub struct StarOffMachine {
        usable: bool,
        dollars: Cell<u64>,
    }

    impl Drop for StarOffMachine {
        fn drop(&mut self) {
            let contents = self.dollars.get();
            println!("Dropping a machine; sending {} dollars to Sylvester.",
                     contents);
            self.dollars.set(0);
            self.usable = false;
        }
    }

    impl StarOffMachine {
        pub fn new() -> StarOffMachine {
            StarOffMachine { usable: true, dollars: Cell::new(0) }
        }
        pub fn remove_star(&self, s: &mut Sneetch) {
            assert!(self.usable,
                    "No different than a read of a dangling pointer.");
            self.dollars.set(self.dollars.get() + 10);
            s.has_star = false;
        }
    }

    pub struct Sneetch<'a> {
        name: &'static str,
        has_star: bool,
        machine: Cell<Option<&'a StarOffMachine>>,
    }

    impl<'a> Sneetch<'a> {
        pub fn new(name: &'static str) -> Sneetch<'a> {
            Sneetch {
                name: name,
                has_star: true,
                machine: Cell::new(None)
            }
        }

        pub fn find_machine(&self, m: &'a StarOffMachine) {
            self.machine.set(Some(m));
        }
    }

    #[unsafe_destructor]
    impl<'a> Drop for Sneetch<'a> {
        fn drop(&mut self) {
            if let Some(m) = self.machine.get() {
                println!("{} says ``before I die, I want to join my \
                          plain-bellied brethren.''", self.name);
                m.remove_star(self);
            }
        }
    }
}

fn unwary_client() {
    use mcbean::{Sneetch, StarOffMachine};
    let (s1, m, s2, s3); // (accommodate PR 21657)
    s1 = Sneetch::new("Sneetch One");
    m = StarOffMachine::new();
    s2 = Sneetch::new("Sneetch Two");
    s3 = Sneetch::new("Sneetch Zee");

    s1.find_machine(&m);
    s2.find_machine(&m);
    s3.find_machine(&m);
}

fn main() {
    unwary_client();
}

This compiles today; if you run it, it prints the following:

Sneetch Zee says ``before I die, I want to join my plain-bellied brethren.''
Sneetch Two says ``before I die, I want to join my plain-bellied brethren.''
Dropping a machine; sending 20 dollars to Sylvester.
Sneetch One says ``before I die, I want to join my plain-bellied brethren.''
thread '<main>' panicked at 'No different than a read of a dangling pointer.', <anon>:27

Explanation: In Sylvester's code, the Drop implementation for Sneetch invokes a method on the borrowed reference in the field machine. This implies there is an implicit restriction on an value s of type Sneetch<'a>: the lifetime 'a must strictly outlive s.

(The example encodes this constraint in a dynamically-checked manner via an explicit usable boolean flag that is only set to false in the machine's own destructor; it is important to keep in mind that this is just a method to illustrate the violation in a semi-reliable manner: Using a machine after usable is set to false by its fn drop code is analogous to dereferencing a *mut T that has been deallocated, or similar soundness violations.)

Sylvester's API does not encode the constraint "'a must strictly outlive the Sneetch<'a>" explicitly; Rust currently has no way of expressing the constraint that one lifetime be strictly greater than another lifetime or type (the form 'a:'b only formally says that 'a must live at least as long as 'b).

Thus, client code like that in unwary_client can inadvertantly set up scenarios where Sylvester's code may break, and Sylvester might be completely unaware of the vulnerability.

Type parameterization: the problem of trait bounds

One might think that all instances of this problem can be identified by the use of a lifetime-parametric Drop implementation, such as impl<'a> Drop for Sneetch<'a> { ..> }

However, consider this trait and struct:

trait Button { fn push(&self); }
struct Zook<B: Button> { button: B, }
#[unsafe_destructor]
impl<B: Button> Drop for Zook<B> {
    fn drop(&mut self) { self.button.push(); }
}

In this case, it is not obvious that there is anything wrong here.

But if we continue the example:

struct Bomb { usable: bool }
impl Drop for Bomb { fn drop(&mut self) { self.usable = false; } }
impl Bomb { fn activate(&self) { assert!(self.usable) } }

enum B<'a> { HarmlessButton, BigRedButton(&'a Bomb) }
impl<'a> Button for B<'a> {
    fn push(&self) {
        if let B::BigRedButton(borrowed) = *self {
            borrowed.activate();
        }
    }
}

fn main() {
    let (mut zook, ticking);
    zook = Zook { button: B::HarmlessButton };
    ticking = Bomb { usable: true };
    zook.button = B::BigRedButton(&ticking);
}

Within the zook there is a hidden reference to borrowed data, ticking, that is assigned the same lifetime as zook but that will be dropped before zook is.

(These examples may seem contrived; see Appendix A for a far less contrived example, that also illustrates how the use of borrowed data can lie hidden behind type parameters.)

The proposal

This RFC is proposes to fix this scenario, by having the compiler ensure that types with destructors are only employed in contexts where either any borrowed data with lifetime 'a within the type either strictly outlives the value of that type, or such borrowed data is provably not accessible from any Drop implementation via a reference of type &'a/&'a mut. This is the "Drop-Check" (aka dropck) rule.

Detailed design

The Drop-Check Rule

The Motivation section alluded to the compiler enforcing a new rule. Here is a more formal statement of that rule:

Let v be some value (either temporary or named) and 'a be some lifetime (scope); if the type of v owns data of type D, where (1.) D has a lifetime- or type-parametric Drop implementation, and (2.) the structure of D can reach a reference of type &'a _, and (3.) either:

  • (A.) the Drop impl for D instantiates D at 'a directly, i.e. D<'a>, or,

  • (B.) the Drop impl for D has some type parameter with a trait bound T where T is a trait that has at least one method,

then 'a must strictly outlive the scope of v.

(Note: This rule is using two phrases that deserve further elaboration and that are discussed further in sections that follow: "the type owns data of type D" and "must strictly outlive".)

(Note: When encountering a D of the form Box<Trait+'b>, we conservatively assume that such a type has a Drop implementation parametric in 'b.)

This rule allows much sound existing code to compile without complaint from rustc. This is largely due to the fact that many Drop implementations enjoy near-complete parametricity: They tend to not impose any bounds at all on their type parameters, and thus the rule does not apply to them.

At the same time, this rule catches the cases where a destructor could possibly reference borrowed data via a reference of type &'a _ or &'a mut_. Here is why:

Condition (A.) ensures that a type like Sneetch<'a> from the Sneetch example will only be assigned to an expression s where 'a strictly outlives s.

Condition (B.) catches cases like Zook<B<'a>> from the Zook example, where the destructor's interaction with borrowed data is hidden behind a method call in the fn drop.

Near-complete parametricity suffices

Noncopy types

All non-Copy type parameters are (still) assumed to have a destructor. Thus, one would be correct in noting that even a type T with no bounds may still have one hidden method attached; namely, its Drop implementation.

However, the drop implementation for T can only be called when running the destructor for value v if either:

  1. the type of v owns data of type T, or

  2. the destructor of v constructs an instance of T.

In the first case, the Drop-Check rule ensures that T must satisfy either Condition (A.) or (B.). In this second case, the freshly constructed instance of T will only be able to access either borrowed data from v itself (and thus such data will already have lifetime that strictly outlives v) or data created during the execution of the destructor.

Any instances

All types implementing Any is forced to outlive 'static. So one should not be able to hide borrowed data behind the Any trait, and therefore it is okay for the analysis to treat Any like a black box whose destructor is safe to run (at least with respect to not accessing borrowed data).

Strictly outlives

There is a notion of "strictly outlives" within the compiler internals. (This RFC is not adding such a notion to the language itself; expressing "'a strictly outlives 'b" as an API constraint is not a strict necessity at this time.)

The heart of the idea is this: we approximate the notion of "strictly outlives" by the following rule: if a value U needs to strictly outlive another value V with code extent S, we could just say that U needs to live at least as long as the parent scope of S.

There are likely to be sound generalizations of the model given here (and we will likely need to consider such to adopt future extensions like Single-Entry-Multiple-Exit (SEME) regions, but that is out of scope for this RFC).

In terms of its impact on the language, the main change has already landed in the compiler; see Rust PR 21657, which added CodeExtent::Remainder, for more direct details on the implications of that change written in a user-oriented fashion.

One important detail of the strictly-outlives relationship that comes in part from Rust PR 21657: All bindings introduced by a single let statement are modeled as having the same lifetime. In an example like

let a;
let b;
let (c, d);
...

a strictly outlives b, and b strictly outlives both c and d. However, c and d are modeled as having the same lifetime; neither one strictly outlives the other. (Of course, during code execution, one of them will be dropped before the other; the point is that when rustc builds its internal model of the lifetimes of data, it approximates and assigns them both the same lifetime.) This is an important detail, because there are situations where one must assign the same lifetime to two distinct bindings in order to allow them to mutually refer to each other's data.

For more details on this "strictly outlives" model, see Appendix B.

When does one type own another

The definition of the Drop-Check Rule used the phrase "if the type owns data of type D".

This criteria is based on recursive descent of the structure of an input type E.

  • If E itself has a Drop implementation that satisfies either condition (A.) or (B.) then add, for all relevant 'a, the constraint that 'a must outlive the scope of the value that caused the recursive descent.

  • Otherwise, if we have previously seen E during the descent then skip it (i.e. we assume a type has no destructor of interest until we see evidence saying otherwise). This check prevents infinite-looping when we encounter recursive references to a type, which can arise in e.g. Option<Box<Type>>.

  • Otherwise, if E is a struct (or tuple), for each of the struct's fields, recurse on the field's type (i.e., a struct owns its fields).

  • Otherwise, if E is an enum, for each of the enum's variants, and for each field of each variant, recurse on the field's type (i.e., an enum owns its fields).

  • Otherwise, if E is of the form & T, &mut T, * T, or fn (T, ...) -> T, then skip this E (i.e., references, native pointers, and bare functions do not own the types they refer to).

  • Otherwise, recurse on any immediate type substructure of E. (i.e., an instantiation of a polymorphic type Poly<T_1, T_2> is assumed to own T_1 and T_2; note that structs and enums do not fall into this category, as they are handled up above; but this does cover cases like Box<Trait<T_1, T_2>+'a>).

Phantom Data

The above definition for type-ownership is (believed to be) sound for pure Rust programs that do not use unsafe, but it does not suffice for several important types without some tweaks.

In particular, consider the implementation of Vec<T>: as of "Rust 1.0 alpha":

pub struct Vec<T> {
    ptr: NonZero<*mut T>,
    len: uint,
    cap: uint,
}

According to the above definition, Vec<T> does not own T. This is clearly wrong.

However, it generalizing the rule to say that *mut T owns T would be too conservative, since there are cases where one wants to use *mut T to model references to state that are not owned.

Therefore, we need some sort of marker, so that types like Vec<T> can express that values of that type own instances of T. The PhantomData<T> marker proposed by RFC 738 ("Support variance for type parameters") is a good match for this. This RFC assumes that either RFC 738 will be accepted, or if necessary, this RFC will be amended so that it itself adds the concept of PhantomData<T> to the language. Therefore, as an additional special case to the criteria above for when the type E owns data of type D, we include:

  • If E is PhantomData<T>, then recurse on T.

Examples of changes imposed by the Drop-Check Rule

Some cyclic structure is still allowed

Earlier versions of the Drop-Check rule were quite conservative, to the point where cyclic data would be disallowed in many contexts. The Drop-Check rule presented in this RFC was crafted to try to keep many existing useful patterns working.

In particular, cyclic structure is still allowed in many contexts. Here is one concrete example:

use std::cell::Cell;

#[derive(Show)]
struct C<'a> {
    v: Vec<Cell<Option<&'a C<'a>>>>,
}

impl<'a> C<'a> {
    fn new() -> C<'a> {
        C { v: Vec::new() }
    }
}

fn f() {
    let (mut c1, mut c2, mut c3);
    c1 = C::new();
    c2 = C::new();
    c3 = C::new();

    c1.v.push(Cell::new(None));
    c1.v.push(Cell::new(None));
    c2.v.push(Cell::new(None));
    c2.v.push(Cell::new(None));
    c3.v.push(Cell::new(None));
    c3.v.push(Cell::new(None));

    c1.v[0].set(Some(&c2));
    c1.v[1].set(Some(&c3));
    c2.v[0].set(Some(&c2));
    c2.v[1].set(Some(&c3));
    c3.v[0].set(Some(&c1));
    c3.v[1].set(Some(&c2));
}

In this code, each of the nodes { c1, c2, c3 } contains a reference to the two other nodes, and those references are stored in a Vec. Note that all of the bindings are introduced by a single let-statement; this is to accommodate the region inference system which wants to assign a single code extent to the 'a lifetime, as discussed in the strictly-outlives section.

Even though Vec<T> itself is defined as implementing Drop, it puts no bounds on T, and therefore that Drop implementation is ignored by the Drop-Check rule.

Directly mixing cycles and Drop is rejected

The Sneetch example illustrates a scenario were borrowed data is dropped while there is still an outstanding borrow that will be accessed by a destructor. In that particular example, one can easily reorder the bindings to ensure that the StarOffMachine outlives all of the sneetches.

But there are other examples that have no such resolution. In particular, graph-structured data where the destructor for each node accesses the neighboring nodes in the graph; this simply cannot be done soundly, because when there are cycles, there is no legal order in which to drop the nodes.

(At least, we cannot do it soundly without imperatively removing a node from the graph as the node is dropped; but we are not going to attempt to support verifying such an invariant as part of this RFC; to my knowledge it is not likely to be feasible with type-checking based static analyses).

In any case, we can easily show some code that will now start to be rejected due to the Drop-Check rule: we take the same C<'a> example of cyclic structure given above, but we now attach a Drop implementation to C<'a>:

use std::cell::Cell;

#[derive(Show)]
struct C<'a> {
    v: Vec<Cell<Option<&'a C<'a>>>>,
}

impl<'a> C<'a> {
    fn new() -> C<'a> {
        C { v: Vec::new() }
    }
}

// (THIS IS NEW)
impl<'a> Drop for C<'a> {
    fn drop(&mut self) { }
}

fn f() {
    let (mut c1, mut c2, mut c3);
    c1 = C::new();
    c2 = C::new();
    c3 = C::new();

    c1.v.push(Cell::new(None));
    c1.v.push(Cell::new(None));
    c2.v.push(Cell::new(None));
    c2.v.push(Cell::new(None));
    c3.v.push(Cell::new(None));
    c3.v.push(Cell::new(None));

    c1.v[0].set(Some(&c2));
    c1.v[1].set(Some(&c3));
    c2.v[0].set(Some(&c2));
    c2.v[1].set(Some(&c3));
    c3.v[0].set(Some(&c1));
    c3.v[1].set(Some(&c2));
}

Now the addition of impl<'a> Drop for C<'a> changes the results entirely;

The Drop-Check rule sees the newly added impl<'a> Drop for C<'a>, which means that for every value of type C<'a>, 'a must strictly outlive the value. But in the binding let (mut c1, mut c2, mut c3) , all three bindings are assigned the same type C<'scope_of_c1_c2_and_c3>, where 'scope_of_c1_c2_and_c3 does not strictly outlive any of the three. Therefore this code will be rejected.

(Note: it is irrelevant that the Drop implementation is a no-op above. The analysis does not care what the contents of that code are; it solely cares about the public API presented by the type to its clients. After all, the Drop implementation for C<'a> could be rewritten tomorrow to contain code that accesses the neighboring nodes.

Some temporaries need to be given names

Due to the way that rustc implements the strictly-outlives relation in terms of code-extents, the analysis does not know in an expression like foo().bar().quux() in what order the temporary values foo() and foo().bar() will be dropped.

Therefore, the Drop-Check rule sometimes forces one to rewrite the code so that it is apparent to the compiler that the value from foo() will definitely outlive the value from foo().bar().

Thus, on occasion one is forced to rewrite:

let q = foo().bar().quux();
...

as:

let foo = foo();
let q = foo.bar().quux()
...

or even sometimes as:

let foo = foo();
let bar = foo.bar();
let q = bar.quux();
...

depending on the types involved.

In practice, pnkfelix saw this arise most often with code like this:

for line in old_io::stdin().lock().lines() {
    ...
}

Here, the result of stdin() is a StdinReader, which holds a RaceBox in a Mutex behind an Arc. The result of the lock() method is a StdinReaderGuard<'a>, which owns a MutexGuard<'a, RaceBox>. The MutexGuard has a Drop implementation that is parametric in 'a; thus, the Drop-Check rule insists that the lifetime assigned to 'a strictly outlive the MutexGuard.

So, under this RFC, we rewrite the code like so:

let stdin = old_io::stdin();
for line in stdin.lock().lines() {
    ...
}

(pnkfelix acknowledges that this rewrite is unfortunate. Potential future work would be to further revise the code extent system so that the compiler knows that the temporary from stdin() will outlive the temporary from stdin().lock(). However, such a change to the code extents could have unexpected fallout, analogous to the fallout that was associated with Rust PR 21657.)

Mixing acyclic structure and Drop is sometimes rejected

This is an example of sound code, accepted today, that is unfortunately rejected by the Drop-Check rule (at least in pnkfelix's prototype):

#![feature(unsafe_destructor)]

use std::cell::Cell;

#[derive(Show)]
struct C<'a> {
    f: Cell<Option<&'a C<'a>>>,
}

impl<'a> C<'a> {
    fn new() -> C<'a> {
        C { f: Cell::new(None), }
    }
}

// force dropck to care about C<'a>
#[unsafe_destructor]
impl<'a> Drop for C<'a> {
    fn drop(&mut self) { }
}

fn f() {
    let c2;
    let mut c1;

    c1 = C::new();
    c2 = C::new();

    c1.f.set(Some(&c2));
}

fn main() {
    f();
}

In principle this should work, since c1 and c2 are assigned to distinct code extents, and c1 will be dropped before c2. However, in the prototype, the region inference system is determining that the lifetime 'a in &'a C<'a> (from the c1.f.set(Some(&c2)); statement) needs to cover the whole block, rather than just the block remainder extent that is actually covered by the let c2;.

(This may just be a bug somewhere in the prototype, but for the time being pnkfelix is going to assume that it will be a bug that this RFC is forced to live with indefinitely.)

Unsound APIs need to be revised or removed entirely

While the Drop-Check rule is designed to ensure that safe Rust code is sound in its use of destructors, it cannot assure us that unsafe code is sound. It is the responsibility of the author of unsafe code to ensure it does not perform unsound actions; thus, we need to audit our own API's to ensure that the standard library is not providing functionality that circumvents the Drop-Check rule.

The most obvious instance of this is the arena crate: in particular: one can use an instance of arena::Arena to create cyclic graph structure where each node's destructor accesses (via &_ references) its neighboring nodes.

Here is a version of our running C<'a> example (where we now do something interesting the destructor for C<'a>) that demonstrates the problem:

Example:

extern crate arena;

use std::cell::Cell;

#[derive(Show)]
struct C<'a> {
    name: &'static str,
    v: Vec<Cell<Option<&'a C<'a>>>>,
    usable: bool,
}

impl<'a> Drop for C<'a> {
    fn drop(&mut self) {
        println!("dropping {}", self.name);
        for neighbor in self.v.iter().map(|v|v.get()) {
            if let Some(neighbor) = neighbor {
                println!("  {} checking neighbor {}",
                         self.name, neighbor.name);
                assert!(neighbor.usable);
            }
        }
        println!("done dropping {}", self.name);
        self.usable = false;

    }
}

impl<'a> C<'a> {
    fn new(name: &'static str) -> C<'a> {
        C { name: name, v: Vec::new(), usable: true }
    }
}

fn f() {
    use arena::Arena;
    let arena = Arena::new();
    let (c1, c2, c3);

    c1 = arena.alloc(|| C::new("c1"));
    c2 = arena.alloc(|| C::new("c2"));
    c3 = arena.alloc(|| C::new("c3"));

    c1.v.push(Cell::new(None));
    c1.v.push(Cell::new(None));
    c2.v.push(Cell::new(None));
    c2.v.push(Cell::new(None));
    c3.v.push(Cell::new(None));
    c3.v.push(Cell::new(None));

    c1.v[0].set(Some(c2));
    c1.v[1].set(Some(c3));
    c2.v[0].set(Some(c2));
    c2.v[1].set(Some(c3));
    c3.v[0].set(Some(c1));
    c3.v[1].set(Some(c2));
}

Calling f() results in the following printout:

dropping c3
  c3 checking neighbor c1
  c3 checking neighbor c2
done dropping c3
dropping c1
  c1 checking neighbor c2
  c1 checking neighbor c3
thread '<main>' panicked at 'assertion failed: neighbor.usable', ../src/test/compile-fail/dropck_untyped_arena_cycle.rs:19

This is unsound. It should not be possible to express such a scenario without using unsafe code.

This RFC suggests that we revise the Arena API by adding a phantom lifetime parameter to its type, and bound the values the arena allocates by that phantom lifetime, like so:

pub struct Arena<'longer_than_self> {
    _invariant: marker::InvariantLifetime<'longer_than_self>,
    ...
}

impl<'longer_than_self> Arena<'longer_than_self> {
    pub fn alloc<T:'longer_than_self, F>(&self, op: F) -> &mut T
        where F: FnOnce() -> T {
        ...
    }
}

Admittedly, this is a severe limitation, since it forces the data allocated by the Arena to store only references to data that strictly outlives the arena, regardless of whether the allocated data itself even has a destructor. (I.e., Arena would become much weaker than TypedArena when attempting to work with cyclic structures). (pnkfelix knows of no way to fix this without adding further extensions to the language, e.g. some way to express "this type's destructor accesses none of its borrowed data", which is out of scope for this RFC.)

Alternatively, we could just deprecate the Arena API, (which is not marked as stable anyway.

The example given here can be adapted to other kinds of backing storage structures, in order to double-check whether the API is likely to be sound or not. For example, the arena::TypedArena<T> type appears to be sound (as long as it carries PhantomData<T> just like Vec<T> does). In particular, when one ports the above example to use TypedArena instead of Arena, it is statically rejected by rustc.

The final goal: remove #[unsafe_destructor]

Once all of the above pieces have landed, lifetime- and type-parameterized Drop will be safe, and thus we will be able to remove #[unsafe_destructor]!

Drawbacks

  • The Drop-Check rule is a little complex, and does disallow some sound code that would compile today.

  • The change proposed in this RFC places restrictions on uses of types with attached destructors, but provides no way for a type Foo<'a> to state as part of its public interface that its drop implementation will not read from any borrowed data of lifetime 'a. (Extending the language with such a feature is potential future work, but is out of scope for this RFC.)

  • Some useful interfaces are going to be disallowed by this RFC. For example, the RFC recommends that the current arena::Arena be revised or simply deprecated, due to its unsoundness. (If desired, we could add an UnsafeArena that continues to support the current Arena API with the caveat that its users need to manually enforce the constraint that the destructors do not access data that has been already dropped. But again, that decision is out of scope for this RFC.)

Alternatives

We considered simpler versions of the Drop-Check rule; in particular, an earlier version of it simply said that if the type of v owns any type D that implements Drop, then for any lifetime 'a that D refers to, 'a must strictly outlive the scope of v, because the destructor for D might hypothetically access borrowed data of lifetime 'a.

  • This rule is simpler in the sense that it more obviously sound.

  • But this rule disallowed far more code; e.g. the Cyclic structure still allowed example was rejected under this more naive rule, because C<'a> owns D = Vec<Cell<Option<&'a C<'a>>>>, and this particular D refers to 'a.


Sticking with the current #[unsafe_destructor] approach to lifetime- and type-parametric types that implement Drop is not really tenable; we need to do something (and we have been planning to do something like this RFC for over a year).

Unresolved questions

  • Is the Drop-Check rule provably sound? pnkfelix has based his argument on informal reasoning about parametricity, but it would be good to put forth a more formal argument. (And in the meantime, pnkfelix invites the reader to try to find holes in the rule, preferably with concrete examples that can be fed into the prototype.)

  • How much can covariance help with some of the lifetime issues?

    See in particular Rust Issue 21198 "new scoping rules for safe dtors may benefit from variance on type params"

Before adding Condition (B.) to the Drop-Check Rule, it seemed like enabling covariance in more standard library types was going to be very important for landing this work. And even now, it is possible that covariance could still play an important role. But nonetheless, there are some API's whose current form is fundamentally incompatible with covariance; e.g. the current TypedArena<T> API is fundamentally invariant with respect to T.

Appendices

Appendix A: Why and when would Drop read from borrowed data

Here is a story, about two developers, Julia and Kurt, and the code they hacked on.

Julia inherited some code, and it is misbehaving. It appears like key/value entries that the code inserts into the standard library's HashMap are not always retrievable from the map. Julia's current hypothesis is that something is causing the keys' computed hash codes to change dynamically, sometime after the entries have been inserted into the map (but it is not obvious when or if this change occurs, nor what its source might be). Julia thinks this hypothesis is plausible, but does not want to audit all of the key variants for possible causes of hash code corruption until after she has hard evidence confirming the hypothesis.

Julia writes some code that walks a hash map's internals and checks that all of the keys produce a hash code that is consistent with their location in the map. However, since it is not clear when the keys' hash codes are changing, it is not clear where in the overall code base she should add such checks. (The hash map is sufficiently large that she cannot simply add calls to do this consistency check everywhere.)

However, there is one spot in the control flow that is a clear contender: if the check is run right before the hash map is dropped, then that would surely be sometime after the hypothesized corruption had occurred. In other words, a destructor for the hash map seems like a good place to start; Julia could make her own local copy of the hash map library and add this check to a impl<K,V,S> Drop for HashMap<K,V,S> { ... } implementation.

In this new destructor code, Julia needs to invoke the hash-code method on K. So she adds the bound where K: Eq + Hash<H> to her HashMap and its Drop implementation, along with the corresponding code to walk the table's entries and check that the hash codes for all the keys matches their position in the table.

Using this, Julia manages confirms her hypothesis (yay). And since it was a reasonable amount of effort to do this experiment, she puts this variation of HashMap up on crates.io, calling it the CheckedHashMap type.

Sometime later, Kurt pulls a copy of CheckHashMap off of crates.io, and he happens to write some code that looks like this:

fn main() {
    #[derive(PartialEq, Eq, Hash, Debug)]
    struct Key<'a> { name: &'a str }

    {
        let (key, mut map, name) : (Key, CheckedHashMap<&Key, String>, String);
        name = format!("k1");
        map = CheckedHashMap::new();
        key = Key { name: &*name };
        map.map.insert(&key, format!("Value for k1"));
    }
}

And, kaboom: when the map goes out of scope, the destructor for CheckedHashMap attempts to compute a hashcode on a reference to key that may not still be valid, and even if key is still valid, it holds a reference to a slice of name that likewise may not still be valid.

This illustrates a case where one might legitimately mix destructor code with borrowed data. (Is this example any less contrived than the Sneetch example? That is in the eye of the beholder.)

Appendix B: strictly-outlives details

The rest of this section gets into some low-level details of parts of how rustc is implemented, largely because the changes described here do have an impact on what results the rustc region inference system produces (or fails to produce). It serves mostly to explain (1.) why Rust PR 21657 was implemented, and (2.) why one may sometimes see indecipherable region-inference errors.

Review: Code Extents

(Nothing here is meant to be new; its just providing context for the next subsection.)

Every Rust expression evaluates to a value V that is either placed into some location with an associated lifetime such as 'l, or V is associated with a block of code that statically delimits the V's runtime extent (i.e. we know from the function's text where V will be dropped). In the rustc source, the blocks of code are sometimes called "scopes" and sometimes "code extents"; I will try to stick to the latter term here, since the word "scope" is terribly overloaded.

Currently, the code extents in Rust are arranged into a tree hierarchy structured similarly to the abstract syntax tree; for any given code extent, the compiler can ask for its parent in this hierarchy.

Every Rust expression E has an associated "terminating extent" somewhere in its chain of parent code extents; temporary values created during the execution of E are stored at stack locations managed by E's terminating extent. When we hit the end of the terminating extent, all such temporaries are dropped.

An example of a terminating extent: in a let-statement like:

let <pat> = <expr>;

the terminating extent of <expr> is the let-statement itself. So in an example like:

let a1 = input.f().g();`
...

there is a temporary value returned from input.f(), and it will live until the end of the let statement, but not into the subsequent code represented by .... (The value resulting from input.f().g(), on the other hand, will be stored in a1 and lives until the end of the block enclosing the let statement.)

(It is not important to this RFC to know the full set of rules dictating which parent expressions are deemed terminating extents; we just will assume that these things do exist.)

For any given code extent S, the parent code extent P of S, if it exists, potentially holds bits of code that will execute after S is done. Any cleanup code for any values assigned to P will only run after we have finished with all code associated with S.

A problem with 1.0 alpha code extents

So, with the above established, we have a hint at how to express that a lifetime 'a needs to strictly outlive a particular code extent S: simply say that 'a needs to live at least long as P.

However, this is a little too simplistic, at least for the Rust compiler circa Rust 1.0 alpha. The main problem is that all the bindings established by let statements in a block are assigned the same code extent.

This, combined with our simplistic definition, yields real problems. For example, in:

{
    use std::fmt;
    #[derive(Debug)] struct DropLoud<T:fmt::Debug>(&'static str, T);
    impl<T:fmt::Debug> Drop for DropLoud<T> {
        fn drop(&mut self) { println!("dropping {}:{:?}", self.0, self.1); }
    }

    let c1 = DropLoud("c1", 1);
    let c2 = DropLoud("c2", &c1);
}

In principle, the code above is legal: c2 will be dropped before c1 is, and thus it is okay that c2 holds a borrowed reference to c1 that will be read when c2 is dropped (indirectly via the fmt::Debug implementation.

However, with the structure of code extents as of Rust 1.0 alpha, c1 and c2 are both given the same code extent: that of the block itself. Thus in that context, this definition of "strictly outlives" indicates that c1 does not strictly outlive c2, because c1 does not live at least as long as the parent of the block; it only lives until the end of the block itself.

This illustrates why "All the bindings established by let statements in a block are assigned the same code extent" is a problem

Block Remainder Code Extents

The solution proposed here (motivated by experience with the prototype) is to introduce finer-grained code extents. This solution is essentially Rust PR 21657, which has already landed in rustc. (That is in part why this is merely an appendix, rather than part of the body of the RFC itself.)

The code extents remain in a tree-hierarchy, but there are now extra entries in the tree, which provide the foundation for a more precise "strictly outlives" relation.

We introduce a new code extent, called a "block remainder" extent, for every let statement in a block, representing the suffix of the block covered by the bindings in that let statement.

For example, given { let (a, b) = EXPR_1; let c = EXPR_2; ... }, which previously had a code extent structure like:

{ let (a, b) = EXPR_1; let c = EXPR_2; ... }
               +----+          +----+
  +------------------+ +-------------+
+------------------------------------------+

so the parent extent of each let statement was the whole block.

But under the new rules, there are two new block remainder extents introduced, with this structure:

{  let (a, b) = EXPR_1;  let c = EXPR_2; ...  }
                +----+           +----+
   +------------------+  +-------------+
                        +-------------------+   <-- new: block remainder 2
  +------------------------------------------+  <-- new: block remainder 1
+---------------------------------------------+

The first let-statement introduces a block remainder extent that covers the lifetime for a and b. The second let-statement introduces a block remainder extent that covers the lifetime for c.

Each let-statement continues to be the terminating extent for its initializer expression. But now, the parent of the extent of the second let statement is a block remainder extent ("block remainder 2"), and, importantly, the parent of block remainder 2 is another block remainder extent ("block remainder 1"). This way, we precisely represent the lifetimes of the named values bound by each let statement, and know that a and b both strictly outlive c as well as the temporary values created during evaluation of EXPR_2. Likewise, c strictly outlives the bindings and temporaries created in the ... that follows it.

Why stop at let-statements?

This RFC does not propose that we attempt to go further and track the order of destruction of the values bound by a single let statement.

Such an experiment could be made part of future work, but for now, we just continue to assign a and b to the same scope; the compiler does not attempt to reason about what order they will be dropped in, and thus we cannot for example reference data borrowed from a in any destructor code for b.

The main reason that we do not want to attempt to produce even finer grain scopes, at least not right now, is that there are scenarios where it is important to be able to assign the same region to two distinct pieces of data; in particular, this often arises when one wants to build cyclic structure, as discussed in Cyclic structure still allowed.