pull/10/head
Alexis Beingessner 10 years ago committed by Manish Goregaokar
parent 5473508096
commit 9d36fdbd0d

@ -0,0 +1,186 @@
% Concurrency and Paralellism
```Not sure if I want this
Safe Rust features *a ton* of tooling to make concurrency and parallelism totally
safe, easy, and fearless. This is a case where we'll really just
[defer to TRPL][trpl-conc] for the basics.
TL;DR: The `Send` and `Sync` traits in conjunction with Rust's ownership model and
normal generic bounds make using concurrent APIs really easy and painless for
a user of Safe Rust.
```
## Data Races and Race Conditions
Safe Rust guarantees an absence of data races, which are defined as:
* two or more threads concurrently accessing a location of memory
* one of them is a write
* one of them is unsynchronized
A data race has Undefined Behaviour, and is therefore impossible to perform
in Safe Rust. Data races are *mostly* prevented through rust's ownership system:
it's impossible to alias a mutable reference, so it's impossible to perform a
data race. Interior mutability makes this more complicated, which is largely why
we have the Send and Sync traits (see below).
However Rust *does not* prevent general race conditions. This is
pretty fundamentally impossible, and probably honestly undesirable. Your hardware
is racy, your OS is racy, the other programs on your computer are racy, and the
world this all runs in is racy. Any system that could genuinely claim to prevent
*all* race conditions would be pretty awful to use, if not just incorrect.
So it's perfectly "fine" for a Safe Rust program to get deadlocked or do
something incredibly stupid with incorrect synchronization. Obviously such a
program isn't very good, but Rust can only hold your hand so far. Still, a
race condition can't violate memory safety in a Rust program on
its own. Only in conjunction with some other unsafe code can a race condition
actually violate memory safety. For instance:
```rust
use std::thread;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::Arc;
let data = vec![1, 2, 3, 4];
// Arc so that the memory the AtomicUsize is stored in still exists for
// the other thread to increment, even if we completely finish executing
// before it. Rust won't compile the program without it, because of the
// lifetime requirements of thread::spawn!
let idx = Arc::new(AtomicUsize::new(0));
let other_idx = idx.clone();
// `move` captures other_idx by-value, moving it into this thread
thread::spawn(move || {
// It's ok to mutate idx because this value
// is an atomic, so it can't cause a Data Race.
other_idx.fetch_add(10, Ordering::SeqCst);
});
// Index with the value loaded from the atomic. This is safe because we
// read the atomic memory only once, and then pass a *copy* of that value
// to the Vec's indexing implementation. This indexing will be correctly
// bounds checked, and there's no chance of the value getting changed
// in the middle. However our program may panic if the thread we spawned
// managed to increment before this ran. A race condition because correct
// program execution (panicing is rarely correct) depends on order of
// thread execution.
println!("{}", data[idx.load(Ordering::SeqCst)]);
if idx.load(Ordering::SeqCst) < data.len() {
unsafe {
// Incorrectly loading the idx *after* we did the bounds check.
// It could have changed. This is a race condition, *and dangerous*
// because we decided to do `get_unchecked`, which is `unsafe`.
println!("{}", data.get_unchecked(idx.load(Ordering::SeqCst)));
}
}
```
## Send and Sync
Not everything obeys inherited mutability, though. Some types allow you to multiply
alias a location in memory while mutating it. Unless these types use synchronization
to manage this access, they are absolutely not thread safe. Rust captures this with
through the `Send` and `Sync` traits.
* A type is Send if it is safe to send it to another thread.
* A type is Sync if it is safe to share between threads (`&T` is Send).
Send and Sync are *very* fundamental to Rust's concurrency story. As such, a
substantial amount of special tooling exists to make them work right. First and
foremost, they're *unsafe traits*. This means that they are unsafe *to implement*,
and other unsafe code can *trust* that they are correctly implemented. Since
they're *marker traits* (they have no associated items like methods), correctly
implemented simply means that they have the intrinsic properties an implementor
should have. Incorrectly implementing Send or Sync can cause Undefined Behaviour.
Send and Sync are also what Rust calls *opt-in builtin traits*.
This means that, unlike every other trait, they are *automatically* derived:
if a type is composed entirely of Send or Sync types, then it is Send or Sync.
Almost all primitives are Send and Sync, and as a consequence pretty much
all types you'll ever interact with are Send and Sync.
Major exceptions include:
* raw pointers are neither Send nor Sync (because they have no safety guards)
* `UnsafeCell` isn't Sync (and therefore `Cell` and `RefCell` aren't)
* `Rc` isn't Send or Sync (because the refcount is shared and unsynchronized)
`Rc` and `UnsafeCell` are very fundamentally not thread-safe: they enable
unsynchronized shared mutable state. However raw pointers are, strictly speaking,
marked as thread-unsafe as more of a *lint*. Doing anything useful
with a raw pointer requires dereferencing it, which is already unsafe. In that
sense, one could argue that it would be "fine" for them to be marked as thread safe.
However it's important that they aren't thread safe to prevent types that
*contain them* from being automatically marked as thread safe. These types have
non-trivial untracked ownership, and it's unlikely that their author was
necessarily thinking hard about thread safety. In the case of Rc, we have a nice
example of a type that contains a `*mut` that is *definitely* not thread safe.
Types that aren't automatically derived can *opt-in* to Send and Sync by simply
implementing them:
```rust
struct MyBox(*mut u8);
unsafe impl Send for MyBox {}
unsafe impl Sync for MyBox {}
```
In the *incredibly rare* case that a type is *inappropriately* automatically
derived to be Send or Sync, then one can also *unimplement* Send and Sync:
```rust
struct SpecialThreadToken(u8);
impl !Send for SpecialThreadToken {}
impl !Sync for SpecialThreadToken {}
```
Note that *in and of itself* it is impossible to incorrectly derive Send and Sync.
Only types that are ascribed special meaning by other unsafe code can possible cause
trouble by being incorrectly Send or Sync.
Most uses of raw pointers should be encapsulated behind a sufficient abstraction
that Send and Sync can be derived. For instance all of Rust's standard
collections are Send and Sync (when they contain Send and Sync types)
in spite of their pervasive use raw pointers to
manage allocations and complex ownership. Similarly, most iterators into these
collections are Send and Sync because they largely behave like an `&` or `&mut`
into the collection.
TODO: better explain what can or can't be Send or Sync. Sufficient to appeal
only to data races?
## Atomics
Rust pretty blatantly just inherits LLVM's model for atomics, which in turn is
largely based off of the C11 model for atomics. This is not due these models
being particularly excellent or easy to understand. Indeed, these models are
quite complex and are known to have several flaws. Rather, it is a pragmatic
concession to the fact that *everyone* is pretty bad at modeling atomics. At very
least, we can benefit from existing tooling and research around C's model.
Trying to fully explain these models is fairly hopeless, so we're just going to
drop that problem in LLVM's lap.
## Actually Doing Things Concurrently
Rust as a language doesn't *really* have an opinion on how to do concurrency or
parallelism. The standard library exposes OS threads and blocking sys-calls
because *everyone* has those and they're uniform enough that you can provide
an abstraction over them in a relatively uncontroversial way. Message passing,
green threads, and async APIs are all diverse enough that any abstraction over
them tends to involve trade-offs that we weren't willing to commit to for 1.0.
However Rust's current design is setup so that you can set up your own
concurrent paradigm or library as you see fit. Just require the right
lifetimes and Send and Sync where appropriate and everything should Just Work
with everyone else's stuff.
[llvm-conc]: http://llvm.org/docs/Atomics.html
[trpl-conc]: https://doc.rust-lang.org/book/concurrency.html

@ -13,18 +13,18 @@ parts and then build a new type out of them. e.g.
```rust
struct Foo {
x: u32,
y: u16,
x: u32,
y: u16,
}
struct Bar {
a: u32,
b: u16,
a: u32,
b: u16,
}
fn reinterpret(foo: Foo) -> Bar {
let Foo { x, y } = foo;
Bar { a: x, b: y }
let Foo { x, y } = foo;
Bar { a: x, b: y }
}
```
@ -57,18 +57,27 @@ but some changes require a cast. These "true casts" are generally regarded as da
problematic actions. True casts revolves around raw pointers and the primitive numeric
types. Here's an exhaustive list of all the true casts:
* rawptr -> rawptr (e.g. `*mut T as *const T` or `*mut T as *mut U`)
* rawptr <-> usize (e.g. `*mut T as usize` or `usize as *mut T`)
* number -> number (e.g. `u32 as i8` or `i16 as f64`)
* c-like enum -> integer/bool (e.g. `DaysOfWeek as u32`)
* `u8` -> `char`
* something about arrays?
TODO: gank the RFC for sweet casts
For number -> number casts, there are quite a few cases to consider:
* unsigned to bigger unsigned will zero-extend losslessly
* unsigned to smaller unsigned will truncate via wrapping
* signed to unsigned will ... TODO rest of this list
* casting between two integers of the same size (e.g. i32 -> u32) is a no-op
* casting from a smaller integer to a bigger integer (e.g. u32 -> u8) will truncate
* casting from a larger integer to a smaller integer (e.g. u8 -> u32) will
* zero-extend if unsigned
* sign-extend if signed
* casting from a float to an integer will round the float towards zero.
* **NOTE: currently this will cause Undefined Behaviour if the rounded
value cannot be represented by the target integer type**. This is a bug
and will be fixed.
* casting from an integer to float will produce the floating point representation
of the integer, rounded if necessary (rounding strategy unspecified).
* casting from an f32 to an f64 is perfect and lossless.
* casting from an f64 to an f32 will produce the closest possible value
(rounding strategy unspecified).
* **NOTE: currently this will cause Undefined Behaviour if the value
is finite but larger or smaller than the largest or smallest finite
value representable by f32**. This is a bug and will be fixed.
The casts involving rawptrs also allow us to completely bypass type-safety
by re-interpretting a pointer of T to a pointer of U for arbitrary types, as

@ -0,0 +1,151 @@
% The Unsafe Rust Programming Language
This document seeks to complement [The Rust Programming Language][] (TRPL).
Where TRPL introduces the language and teaches the basics, TURPL dives deep into
the specification of the language, and all the nasty bits necessary to write
Unsafe Rust. TURPL does not assume you have read TRPL, but does assume you know
the basics of the language and systems programming. We will not explain the
stack or heap, we will not explain the syntax.
## A Tale Of Two Languages
Rust can be thought of as two different languages: Safe Rust, and Unsafe Rust.
Any time someone opines the guarantees of Rust, they are almost surely talking about
Safe Rust. However Safe Rust is not sufficient to write every program. For that,
we need the Unsafe Rust superset.
Most fundamentally, writing bindings to other languages
(such as the C exposed by your operating system) is never going to be safe. Rust
can't control what other languages do to program execution! However Unsafe Rust is
also necessary to construct fundamental abstractions where the type system is not
sufficient to automatically prove what you're doing is sound.
Indeed, the Rust standard library is implemented in Rust, and it makes substantial
use of Unsafe Rust for implementing IO, memory allocation, collections,
synchronization, and other low-level computational primitives.
Upon hearing this, many wonder why they would not simply just use C or C++ in place of
Rust (or just use a "real" safe language). If we're going to do unsafe things, why not
lean on these much more established languages?
The most important difference between C++ and Rust is a matter of defaults:
Rust is 100% safe by default. Even when you *opt out* of safety in Rust, it is a modular
action. In deciding to work with unchecked uninitialized memory, this does not
suddenly make dangling or null pointers a problem. When using unchecked indexing on `x`,
one does not have to suddenly worry about indexing out of bounds on `y`.
C and C++, by contrast, have pervasive unsafety baked into the language. Even the
modern best practices like `unique_ptr` have various safety pitfalls.
It should also be noted that writing Unsafe Rust should be regarded as an exceptional
action. Unsafe Rust is often the domain of *fundamental libraries*. Anything that needs
to make FFI bindings or define core abstractions. These fundamental libraries then expose
a *safe* interface for intermediate libraries and applications to build upon. And these
safe interfaces make an important promise: if your application segfaults, it's not your
fault. *They* have a bug.
And really, how is that different from *any* safe language? Python, Ruby, and Java libraries
can internally do all sorts of nasty things. The languages themselves are no
different. Safe languages regularly have bugs that cause critical vulnerabilities.
The fact that Rust is written with a healthy spoonful of Unsafe Rust is no different.
However it *does* mean that Rust doesn't need to fall back to the pervasive unsafety of
C to do the nasty things that need to get done.
## What does `unsafe` mean?
Rust tries to model memory safety through the `unsafe` keyword. Interestingly,
the meaning of `unsafe` largely revolves around what
its *absence* means. If the `unsafe` keyword is absent from a program, it should
not be possible to violate memory safety under *any* conditions. The presence
of `unsafe` means that there are conditions under which this code *could*
violate memory safety.
To be more concrete, Rust cares about preventing the following things:
* Dereferencing null/dangling pointers
* Reading uninitialized memory
* Breaking the pointer aliasing rules (TBD) (llvm rules + noalias on &mut and & w/o UnsafeCell)
* Invoking Undefined Behaviour (in e.g. compiler intrinsics)
* Producing invalid primitive values:
* dangling/null references
* a `bool` that isn't 0 or 1
* an undefined `enum` discriminant
* a `char` larger than char::MAX
* A non-utf8 `str`
* Unwinding into an FFI function
* Causing a data race
However libraries are free to declare arbitrary requirements if they could transitively
cause memory safety issues. However Rust is otherwise quite permisive with respect to
other dubious operations. Rust considers it "safe" to:
* Deadlock
* Leak memory
* Fail to call destructors
* Access private fields
* Overflow integers
* Delete the production database
However any program that does such a thing is *probably* incorrect. Rust just isn't
interested in modeling these problems, as they are much harder to prevent in general,
and it's basically impossible to prevent incorrect programs from getting written.
Their are several places `unsafe` can appear in Rust today, which can largely be
grouped into two categories:
* There are unchecked contracts here. To declare you understand this, I require
you to write `unsafe` elsewhere:
* On functions, `unsafe` is declaring the function to be unsafe to call. Users
of the function must check the documentation to determine what this means,
and then have to write `unsafe` somewhere to identify that they're aware of
the danger.
* On trait declarations, `unsafe` is declaring that *implementing* the trait
is an unsafe operation, as it has contracts that other unsafe code is free to
trust blindly.
* I am declaring that I have, to the best of my knowledge, adhered to the
unchecked contracts:
* On trait implementations, `unsafe` is declaring that the contract of the
`unsafe` trait has been upheld.
* On blocks, `unsafe` is declaring any unsafety from an unsafe
operation to be handled, and therefore the parent function is safe.
There is also `#[unsafe_no_drop_flag]`, which is a special case that exists for
historical reasons and is in the process of being phased out. See the section on
destructors for details.
Some examples of unsafe functions:
* `slice::get_unchecked` will perform unchecked indexing, allowing memory
safety to be freely violated.
* `ptr::offset` in an intrinsic that invokes Undefined Behaviour if it is
not "in bounds" as defined by LLVM (see the lifetimes section for details).
* `mem::transmute` reinterprets some value as having the given type,
bypassing type safety in arbitrary ways. (see the conversions section for details)
* All FFI functions are `unsafe` because they can do arbitrary things.
C being an obvious culprit, but generally any language can do something
that Rust isn't happy about. (see the FFI section for details)
As of Rust 1.0 there are exactly two unsafe traits:
* `Send` is a marker trait (it has no actual API) that promises implementors
are safe to send to another thread.
* `Sync` is a marker trait that promises that threads can safely share
implementors through a shared reference.
All other traits that declare any kind of contract *really* can't be trusted
to adhere to their contract when memory-safety is at stake. For instance Rust has
`PartialOrd` and `Ord` to differentiate between types which can "just" be
compared and those that implement a total ordering. However you can't actually
trust an implementor of `Ord` to actually provide a total ordering if failing to
do so causes you to e.g. index out of bounds. But if it just makes your program
do a stupid thing, then it's "fine" to rely on `Ord`.
The reason this is the case is that `Ord` is safe to implement, and it should be
impossible for bad *safe* code to violate memory safety. Rust has traditionally
avoided making traits unsafe because it makes `unsafe` pervasive in the language,
which is not desirable. The only reason `Send` and `Sync` are unsafe is because
thread safety is a sort of fundamental thing that a program can't really guard
against locally (even by-value message passing still requires a notion Send).

@ -1,13 +1,467 @@
% Advanced Lifetimes
% Ownership
Lifetimes are the breakout feature of Rust.
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.
# Safe Rust
* no aliasing of &mut
# Unsafe Rust
## 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:
```rust
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.
## 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
Almost always, the mutability of a lifetime can be derived from the mutability
of the reference it is attached to. However this is not necessarily the case.
For instance in the following code:
```rust
fn foo<'a>(input: &'a mut u8) -> &'a u8 { &* input }
```
One would expect the output of foo to be an immutable lifetime. However we have
derived it from the input, which is a mutable lifetime. So although we have a
shared reference, it will have the much more limited aliasing rules of a mutable
reference. As a consequence, there is no expressive benefit in a method that
mutates returning a shared reference.
## 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:
```rust
&'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:
```rust
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 route
is to just use a small function to ensure the lifetime is bound. 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.
## 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.
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:
```rust
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:
```rust
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 that *takes* a function `len` that takes a function `F`.
```rust
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 print 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.
## PhantomData and PhantomFn
This is all well and good for the types the standard library provides, but
how is variance determined for type that *you* define? The variance of a type
over its generic arguments is determined by how they're stored.
```rust
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
}
```
However 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
reachable. 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.
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>,
}
```
## 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:
```rust
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:
```rust
let x = [1, 2, 3];
let a = &mut x[0];
let b = &mut x[1];
println!("{} {}", a, b);
```
```text
<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:
```rust
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
* Splitting lifetimes into disjoint regions
* Creating lifetimes from raw pointers
*
Loading…
Cancel
Save