rust/lifetimes.md
Alexis Beingessner bb32315226 fix paper link
2015-06-21 09:57:44 -07:00

24 KiB

% Ownership and Lifetimes

Ownership is the breakout feature of Rust. It allows Rust to be completely memory-safe and efficient, while avoiding garbage collection. Before getting into the ownership system in detail, we will consider a simple but fundamental language-design problem.

The Tagged Union Problem

The core of the lifetime and mutability system derives from a simple problem: internal pointers to tagged unions. For instance, consider the following code:

enum Foo {
    A(u32),
    B(f64),
}

let mut x = B(2.0);
if let B(ref mut y) = x {
    *x = A(7);
    // OH NO! a u32 has been interpretted as an f64! Type-safety hole!
    // (this does not actually compile)
    println!("{}", y);

}

The problem here is an intersection of 3 choices:

  • data in a tagged union is inline with the tag
  • tagged unions are mutable
  • being able to take a pointer into a tagged union

Remove any of these 3 and the problem goes away. Traditionally, functional languages have avoided this problem by removing the mutable option. This means that they can in principle keep their data inline (ghc has a pragma for this). A garbage collected imperative language like Java could alternatively solve this problem by just keeping all variants elsewhere, so that changing the variant of a tagged union just overwrites a pointer, and anyone with an outstanding pointer to the inner data is unaffected thanks to The Magic Of Garbage Collection.

Rust, by contrast, takes a subtler approach. Rust allows mutation, allows pointers to inner data, and its enums have their data allocated inline. However it prevents anything from being mutated while there are outstanding pointers to it! And this is all done at compile time.

Interestingly, Rust's std::cell module exposes two types that offer an alternative approach to this problem:

  • The Cell type allows mutation of aliased data, but instead forbids internal pointers to that data. The only way to read or write a Cell is to copy the bits in or out.

  • The RefCell type allows mutation of aliased data and internal pointers, but manages this through runtime checks. It is effectively a thread-unsafe read-write lock.

For more details see Dan Grossman's Existential Types for Imperative Languages:

Lifetimes

Rust's static checks are managed by the borrow checker (borrowck), which tracks mutability and outstanding loans. This analysis can in principle be done without any help locally. However as soon as data starts crossing the function boundary, we have some serious trouble. In principle, borrowck could be a massive whole-program analysis engine to handle this problem, but this would be an atrocious solution. It would be terribly slow, and errors would be horribly non-local.

Instead, Rust tracks ownership through lifetimes. Every single reference and value in Rust is tagged with a lifetime that indicates the scope it is valid for. Rust has two kinds of reference:

  • Shared reference: &
  • Mutable reference: &mut

The main rules are as follows:

  • A shared reference can be aliased
  • A mutable reference cannot be aliased
  • A reference cannot outlive its referrent (&'a T -> T: 'a)

However non-mutable variables have some special rules:

  • You cannot mutate or mutably borrow a non-mut variable,

Only variables marked as mutable can be borrowed mutably, though this is little more than a local lint against incorrect usage of a value.

Weird Lifetimes

Given the following code:

struct Foo;

impl Foo {
    fn mutate_and_share(&mut self) -> &Self { &*self }
    fn share(&self) {}
}

fn main() {
    let mut foo = Foo;
    let loan = foo.mutate_and_share();
    foo.share();
}

One might expect it to compile. We call mutate_and_share, which mutably borrows foo temporarily, but then returns only a shared reference. Therefore we would expect foo.share() to succeed as foo shouldn't be mutably borrowed.

However when we try to compile it:

<anon>:11:5: 11:8 error: cannot borrow `foo` as immutable because it is also borrowed as mutable
<anon>:11     foo.share();
              ^~~
<anon>:10:16: 10:19 note: previous borrow of `foo` occurs here; the mutable borrow prevents subsequent moves, borrows, or modification of `foo` until the borrow ends
<anon>:10     let loan = foo.mutate_and_share();
                         ^~~
<anon>:12:2: 12:2 note: previous borrow ends here
<anon>:8 fn main() {
<anon>:9     let mut foo = Foo;
<anon>:10     let loan = foo.mutate_and_share();
<anon>:11     foo.share();
<anon>:12 }
          ^

What happened? Well, the lifetime of loan is derived from a mutable borrow. This makes the type system believe that foo is mutably borrowed as long as loan exists, even though it's a shared reference. To my knowledge, this is not a bug.

Lifetime Elision

In order to make common patterns more ergonomic, Rust allows lifetimes to be elided in function, impl, and type signatures.

A lifetime position is anywhere you can write a lifetime in a type:

&'a T
&'a mut T
T<'a>

Lifetime positions can appear as either "input" or "output":

  • For fn definitions, input refers to the types of the formal arguments in the fn definition, while output refers to result types. So fn foo(s: &str) -> (&str, &str) has elided one lifetime in input position and two lifetimes in output position. Note that the input positions of a fn method definition do not include the lifetimes that occur in the method's impl header (nor lifetimes that occur in the trait header, for a default method).

  • In the future, it should be possible to elide impl headers in the same manner.

Elision rules are as follows:

  • Each elided lifetime in input position becomes a distinct lifetime parameter.

  • If there is exactly one input lifetime position (elided or not), that lifetime is assigned to all elided output lifetimes.

  • If there are multiple input lifetime positions, but one of them is &self or &mut self, the lifetime of self is assigned to all elided output lifetimes.

  • Otherwise, it is an error to elide an output lifetime.

Examples:

fn print(s: &str);                                      // elided
fn print<'a>(s: &'a str);                               // expanded

fn debug(lvl: uint, s: &str);                           // elided
fn debug<'a>(lvl: uint, s: &'a str);                    // expanded

fn substr(s: &str, until: uint) -> &str;                // elided
fn substr<'a>(s: &'a str, until: uint) -> &'a str;      // expanded

fn get_str() -> &str;                                   // ILLEGAL

fn frob(s: &str, t: &str) -> &str;                      // ILLEGAL

fn get_mut(&mut self) -> &mut T;                        // elided
fn get_mut<'a>(&'a mut self) -> &'a mut T;              // expanded

fn args<T:ToCStr>(&mut self, args: &[T]) -> &mut Command                  // elided
fn args<'a, 'b, T:ToCStr>(&'a mut self, args: &'b [T]) -> &'a mut Command // expanded

fn new(buf: &mut [u8]) -> BufWriter;                    // elided
fn new<'a>(buf: &'a mut [u8]) -> BufWriter<'a>          // expanded

Unbounded Lifetimes

Unsafe code can often end up producing references or lifetimes out of thin air. Such lifetimes come into the world as unbounded. The most common source of this is derefencing a raw pointer, which produces a reference with an unbounded lifetime. Such a lifetime becomes as big as context demands. This is in fact more powerful than simply becoming 'static, because for instance &'static &'a T will fail to typecheck, but the unbound lifetime will perfectly mold into &'a &'a T as needed. However for most intents and purposes, such an unbounded lifetime can be regarded as 'static.

Almost no reference is 'static, so this is probably wrong. transmute and transmute_copy are the two other primary offenders. One should endeavour to bound an unbounded lifetime as quick as possible, especially across function boundaries.

Given a function, any output lifetimes that don't derive from inputs are unbounded. For instance:

fn get_str<'a>() -> &'a str;

will produce an &str with an unbounded lifetime. The easiest way to avoid unbounded lifetimes is to use lifetime elision at the function boundary. If an output lifetime is elided, then it must be bounded by an input lifetime. Of course, it might be bounded by the wrong lifetime, but this will usually just cause a compiler error, rather than allow memory safety to be trivially violated.

Within a function, bounding lifetimes is more error-prone. The safest and easiest way to bound a lifetime is to return it from a function with a bound lifetime. However if this is unacceptable, the reference can be placed in a location with a specific lifetime. Unfortunately it's impossible to name all lifetimes involved in a function. To get around this, you can in principle use copy_lifetime, though these are unstable due to their awkward nature and questionable utility.

Higher-Rank Lifetimes

Generics in Rust generally allow types to be instantiated with arbitrary associated lifetimes, but this fixes the lifetimes they work with once instantiated. For almost all types, this is exactly the desired behaviour. For example slice::Iter can work with arbitrary lifetimes, determined by the slice that instantiates it. However once Iter is instantiated the lifetimes it works with cannot be changed. It returns references that live for some particular 'a.

However some types are more flexible than this. In particular, a single instantiation of a function can process arbitrary lifetimes:

fn identity(input: &u8) -> &u8 { input }

What is the lifetime that identity works with? There is none. If you think this is "cheating" because functions are statically instantiated, then you need only consider the equivalent closure:

let identity = |input: &u8| input;

These functions are higher ranked over the lifetimes they work with. This means that they're generic over what they handle after instantiation. For most things this would pose a massive problem, but because lifetimes don't exist at runtime, this is really just a compile-time mechanism. The Fn traits contain sugar that allows higher-rank lifetimes to simply be expressed by simply omitting lifetimes:

fn main() {
    foo(|input| input);
}

fn foo<F>(f: F)
    // F is higher-ranked over the lifetime these references have
    where F: Fn(&u8) -> &u8
{
    f(&0);
    f(&1);
}

The desugaring of this is actually unstable:

#![feature(unboxed_closures)]

fn main() {
    foo(|input| input);
}

fn foo<F>(f: F)
    where F: for<'a> Fn<(&'a u8,), Output=&'a u8>
{
    f(&0);
    f(&1);
}

for<'a> is how we declare a higher-ranked lifetime. Unfortunately higher-ranked lifetimes are still fairly new, and are missing a few features to make them maximally useful outside of the Fn traits.

Subtyping and Variance

Although Rust doesn't have any notion of inheritance, it does include subtyping. In Rust, subtyping derives entirely from lifetimes. Since lifetimes are derived from scopes, we can partially order them based on an outlives relationship. We can even express this as a generic bound: T: 'a specifies that T outlives 'a.

We can then define subtyping on lifetimes in terms of lifetimes: 'a : 'b implies 'a <: b -- if 'a outlives 'b, then 'a is a subtype of 'b. This is a very large source of confusion, because a bigger scope is a sub type of a smaller scope. This does in fact make sense. The intuitive reason for this is that if you expect an &'a u8, then it's totally fine for me to hand you an &'static u8, in the same way that if you expect an Animal in Java, it's totally fine for me to hand you a Cat.

(Note, the subtyping relationship and typed-ness of lifetimes is a fairly arbitrary construct that some disagree with. I just find that it simplifies this analysis.)

Variance is where things get really harsh.

Variance is a property that type constructors have. A type constructor in Rust is a generic type with unbound arguments. For instance Vec is a type constructor that takes a T and returns a Vec<T>. & and &mut are type constructors that take a lifetime and a type.

A type constructor's variance is how the subtypes of its inputs affects the subtypes of its outputs. There are three kinds of variance:

  • F is covariant if T <: U implies F<T> <: F<U>
  • F is contravariant if T <: U implies F<U> <: F<T>
  • F is invariant otherwise (no subtyping relation can be derived)

Some important variances:

  • & is covariant (as is *const by metaphor)
  • &mut is invariant (as is *mut by metaphor)
  • Fn(T) is contravariant with respect to T
  • Box, Vec, and all other collections are covariant
  • UnsafeCell, Cell, RefCell, Mutex and all "interior mutability" types are invariant

To understand why these variances are correct and desirable, we will consider several examples. We have already covered why & should be covariant.

To see why &mut should be invariant, consider the following code:

fn main() {
    let mut forever_str: &'static str = "hello";
    {
        let string = String::from("world");
        overwrite(&mut forever_str, &mut &*string);
    }
    println!("{}", forever_str);
}

fn overwrite<T: Copy>(input: &mut T, new: &mut T) {
    *input = *new;
}

The signature of overwrite is clearly valid: it takes mutable references to two values of the same type, and replaces one with the other. We have seen already that & is covariant, and 'static is a subtype of any 'a, so &'static str is a subtype of &'a str. Therefore, if &mut was also covariant, then the lifetime of the &'static str would successfully be "shrunk" down to the shorter lifetime of the string, and replace would be called successfully. The string would subsequently be dropped, and forever_str would point to freed memory when we print it!

Therefore &mut should be invariant. This is the general theme of covariance vs invariance: if covariance would allow you to store a short-lived value in a longer-lived slot, then you must be invariant.

Box and Vec are interesting cases because they're covariant, but you can definitely store values in them! This is fine because you can only store values in them through a mutable reference! The mutable reference makes the whole type invariant, and therefore prevents you from getting in trouble.

Being covariant allows them to be covariant when shared immutably (so you can pass a &Box<&'static str> where a &Box<&'a str> is expected). It also allows you to forever weaken the type by moving it into a weaker slot. That is, you can do:

fn get_box<'a>(&'a u8) -> Box<&'a str> {
    Box::new("hello")
}

which is fine because unlike the mutable borrow case, there's no one else who "remembers" the old lifetime in the box.

The variance of the cell types similarly follows. & is like an &mut for a cell, because you can still store values in them through an &. Therefore cells must be invariant to avoid lifetime smuggling.

Fn is the most confusing case, largely because contravariance is easily the most confusing kind of variance, and basically never comes up. To understand it, consider a function len that takes a function F.

fn len<F>(func: F) -> usize
    where F: Fn(&'static str) -> usize
{
    func("hello")
}

We require that F is a Fn that can take an &'static str and returns a usize. Now say we have a function that can take an &'a str (for some 'a). Such a function actually accepts more inputs, since &'static str is a subtype of &'a str. Therefore len should happily accept such a function!

So a Fn(&'a str) is a subtype of a Fn(&'static str) because &'static str is a subtype of &'a str. Exactly contravariance.

The variance of *const and *mut is basically arbitrary as they're not at all type or memory safe, so their variance is determined in analogy to & and &mut respectively.

This is all well and good for the types the standard library provides, but how is variance determined for type that you define? A struct informally speaking inherits the variance of its fields. If a struct Foo has a generic argument A that is used in a field a, then Foo's variance over A is exactly a's variance. However this is complicated if A is used in multiple fields.

  • If all uses of A are covariant, then Foo is covariant over A
  • If all uses of A are contravariant, then Foo is contravariant over A
  • Otherwise, Foo is invariant over A
struct Foo<'a, 'b, A, B, C, D, E, F, G, H> {
    a: &'a A,     // covariant over 'a and A
    b: &'b mut B, // invariant over 'b and B
    c: *const C,  // covariant over C
    d: *mut D,    // invariant over D
    e: Vec<E>,    // covariant over E
    f: Cell<F>,   // invariant over F
    g: G          // covariant over G
    h1: H         // would also be covariant over H except...
    h2: Cell<H>   // invariant over H, because invariance wins
}

PhantomData

When working with unsafe code, we can often end up in a situation where types or lifetimes are logically associated with a struct, but not actually part of a field. This most commonly occurs with lifetimes. For instance, the Iter for &'a [T] is (approximately) defined as follows:

pub struct Iter<'a, T: 'a> {
    ptr: *const T,
    end: *const T,
}

However because 'a is unused within the struct's body, it's unbound. Because of the troubles this has historically caused, unbound lifetimes and types are illegal in struct definitions. Therefore we must somehow refer to these types in the body. Correctly doing this is necessary to have correct variance and drop checking.

We do this using PhantomData, which is a special marker type. PhantomData consumes no space, but simulates a field of the given type for the purpose of variance. This was deemed to be less error-prone than explicitly telling the type-system the kind of variance that you want.

Iter logically contains &'a T, so this is exactly what we tell the PhantomData to simulate:

pub struct Iter<'a, T: 'a> {
    ptr: *const T,
    end: *const T,
    _marker: marker::PhantomData<&'a T>,
}

Dropck

When a type is going out of scope, Rust will try to Drop it. Drop executes arbitrary code, and in fact allows us to "smuggle" arbitrary code execution into many places. As such additional soundness checks (dropck) are necessary to ensure that a type T can be safely instantiated and dropped. It turns out that we really don't need to care about dropck in practice, as it often "just works".

However the one exception is with PhantomData. Given a struct like Vec:

struct Vec<T> {
    data: *const T, // *const for covariance!
    len: usize,
    cap: usize,
}

dropck will generously determine that Vec does not contain any values of type T. This will unfortunately allow people to construct unsound Drop implementations that access data that has already been dropped. In order to tell dropck that we do own values of type T and may call destructors of that type, we must add extra PhantomData:

struct Vec<T> {
    data: *const T, // *const for covariance!
    len: usize,
    cap: usize,
    _marker: marker::PhantomData<T>,
}

Raw pointers that own an allocation is such a pervasive pattern that the standard library made a utility for itself called Unique<T> which:

  • wraps a *const T,
  • includes a PhantomData,
  • auto-derives Send/Sync as if T was contained
  • marks the pointer as NonZero for the null-pointer optimization

Splitting Lifetimes

The mutual exclusion property of mutable references can be very limiting when working with a composite structure. Borrowck understands some basic stuff, but will fall over pretty easily. Borrowck understands structs sufficiently to understand that it's possible to borrow disjoint fields of a struct simultaneously. So this works today:

struct Foo {
    a: i32,
    b: i32,
    c: i32,
}

let mut x = Foo {a: 0, b: 0, c: 0};
let a = &mut x.a;
let b = &mut x.b;
let c = &x.c;
*b += 1;
let c2 = &x.c;
*a += 10;
println!("{} {} {} {}", a, b, c, c2);

However borrowck doesn't understand arrays or slices in any way, so this doesn't work:

let x = [1, 2, 3];
let a = &mut x[0];
let b = &mut x[1];
println!("{} {}", a, b);
<anon>:3:18: 3:22 error: cannot borrow immutable indexed content `x[..]` as mutable
<anon>:3     let a = &mut x[0];
                          ^~~~
<anon>:4:18: 4:22 error: cannot borrow immutable indexed content `x[..]` as mutable
<anon>:4     let b = &mut x[1];
                          ^~~~
error: aborting due to 2 previous errors

While it was plausible that borrowck could understand this simple case, it's pretty clearly hopeless for borrowck to understand disjointness in general container types like a tree, especially if distinct keys actually do map to the same value.

In order to "teach" borrowck that what we're doing is ok, we need to drop down to unsafe code. For instance, mutable slices expose a split_at_mut function that consumes the slice and returns two mutable slices. One for everything to the left of the index, and one for everything to the right. Intuitively we know this is safe because the slices don't alias. However the implementation requires some unsafety:

fn split_at_mut(&mut self, mid: usize) -> (&mut [T], &mut [T]) {
    unsafe {
        let self2: &mut [T] = mem::transmute_copy(&self);

        (ops::IndexMut::index_mut(self, ops::RangeTo { end: mid } ),
         ops::IndexMut::index_mut(self2, ops::RangeFrom { start: mid } ))
    }
}

This is pretty plainly dangerous. We use transmute to duplicate the slice with an unbounded lifetime, so that it can be treated as disjoint from the other until we unify them when we return.

However more subtle is how iterators that yield mutable references work. The iterator trait is defined as follows:

trait Iterator {
    type Item;

    fn next(&mut self) -> Option<Self::Item>;
}

Given this definition, Self::Item has no connection to self. This means that we can call next several times in a row, and hold onto all the results concurrently. This is perfectly fine for by-value iterators, which have exactly these semantics. It's also actually fine for shared references, as they admit arbitrarily many references to the same thing (although the iterator needs to be a separate object from the thing being shared). But mutable references make this a mess. At first glance, they might seem completely incompatible with this API, as it would produce multiple mutable references to the same object!

However it actually does work, exactly because iterators are one-shot objects. Everything an IterMut yields will be yielded at most once, so we don't actually ever yield multiple mutable references to the same piece of data.

In general all mutable iterators require some unsafe code somewhere, though. Whether it's raw pointers, or safely composing on top of another IterMut.

For instance, VecDeque's IterMut:

pub struct IterMut<'a, T:'a> {
    // The whole backing array. Some of these indices are initialized!
    ring: &'a mut [T],
    tail: usize,
    head: usize,
}

impl<'a, T> Iterator for IterMut<'a, T> {
    type Item = &'a mut T;

    fn next(&mut self) -> Option<&'a mut T> {
        if self.tail == self.head {
            return None;
        }
        let tail = self.tail;
        self.tail = wrap_index(self.tail.wrapping_add(1), self.ring.len());

        unsafe {
            // might as well do unchecked indexing since wrap_index has us
            // in-bounds, and many of the "middle" indices are uninitialized
            // anyway.
            let elem = self.ring.get_unchecked_mut(tail);

            // round-trip through a raw pointer to unbound the lifetime from
            // ourselves
            Some(&mut *(elem as *mut _))
        }
    }
}

A very subtle but interesting detail in this design is that it relies on privacy to be sound. Borrowck works on some very simple rules. One of those rules is that if we have a live &mut Foo and Foo contains an &mut Bar, then that &mut Bar is also live. Since IterMut is always live when next can be called, if ring were public then we could mutate ring while outstanding mutable borrows to it exist!