Self-referencial structs using async stacks
Escher is an extremely simple library providing a safe and sound API to build self-referencial structs. It works by (ab)using the async await trasformation of rustc. If you'd like to know more about the inner workings please take a look at the How it works section and the source code.
Compared to the state of the art escher:
- Is only around 100 lines of well-commented code
- Contains only two
unsafe
calls that are well argued for - Uses rustc for all the analysis. If it compiles, the self references are correct
This library provides the Escher wrapper type that can hold self-referencial data and expose them safely through the as_ref() and as_mut() functions.
You construct a self reference by calling Escher's constructor and providing an async closure
that will initialize your self-references on its stack. Your closure will be provided with a
capturer r
that has a single capture() method that consumes r
.
Note: It is important to
.await
the result.capture()
in order for escher to correctly initialize your struct.
Once all the data and references are created you can capture the desired ones. Simple references to owned data can be captured directly (see first example).
To capture more than one variable or capture references to non-owned data you will have to define your own reference struct that derives Rebindable (see second example).
The simplest way to use Escher is to create a reference of some data and then capture it:
use escher::Escher;
let escher_heart = Escher::new(|r| async move {
let data: Vec<u8> = vec![240, 159, 146, 150];
let sparkle_heart = std::str::from_utf8(&data).unwrap();
r.capture(sparkle_heart).await;
});
assert_eq!("π", *escher_heart.as_ref());
In order to capture more than one things you can define a struct that will be used to capture the variables:
use escher::{Escher, Rebindable};
#[derive(Rebindable)]
struct VecStr<'this> {
data: &'this Vec<u8>,
s: &'this str,
}
let escher_heart = Escher::new(|r| async move {
let data: Vec<u8> = vec![240, 159, 146, 150];
r.capture(VecStr{
data: &data,
s: std::str::from_utf8(&data).unwrap(),
}).await;
});
assert_eq!(240, escher_heart.as_ref().data[0]);
assert_eq!("π", escher_heart.as_ref().s);
If you capture a mutable reference to some piece of data then you cannot capture the data itself like the previous example. This is mandatory as doing otherwise would create two mutable references into the same piece of data which is not allowed.
use escher::Escher;
let mut name = Escher::new(|r| async move {
let mut data: Vec<u8> = vec![101, 115, 99, 104, 101, 114];
let name = std::str::from_utf8_mut(&mut data).unwrap();
r.capture(name).await;
});
assert_eq!("escher", *name.as_ref());
name.as_mut().make_ascii_uppercase();
assert_eq!("ESCHER", *name.as_ref());
use escher::{Escher, Rebindable};
#[derive(Rebindable)]
struct MyStruct<'this> {
int_data: &'this Box<i32>,
int_ref: &'this i32,
float_ref: &'this mut f32,
}
let mut my_value = Escher::new(|r| async move {
let int_data = Box::new(42);
let mut float_data = Box::new(3.14);
r.capture(MyStruct{
int_data: &int_data,
int_ref: &int_data,
float_ref: &mut float_data,
}).await;
});
assert_eq!(Box::new(42), *my_value.as_ref().int_data);
assert_eq!(3.14, *my_value.as_ref().float_ref);
*my_value.as_mut().float_ref = (*my_value.as_ref().int_ref as f32) * 2.0;
assert_eq!(84.0, *my_value.as_ref().float_ref);
The main problem with self-referencial structs is that if such a struct was somehow constructed the compiler would then have to statically prove that it would not move again. This analysis is necessary because any move would invalidate self-pointers since all pointers in rust are absolute memory addresses.
To illustrate why this is necessary, imagine we define a self-referencial struct that holds a Vec and a pointer to it at the same time:
struct Foo {
s: Vec<u8>,
p: &Vec<u8>,
}
Then, let's assume we had a way of getting an instance of this struct. We could then write the following code that creates a dangling pointer in safe rust!
let foo = Foo::magic_construct();
let bar = foo; // move foo to a new location
println!("{:?}", bar.p); // access the self-reference, memory error!
While rust doesn't allow you to explicitly write out self referencial struct members and initialize them it is perfectly valid to write out the values of the members individually as separate stack bindings. This is because the borrow checker can do a move analysis when the values are on the stack.
Practically, we could convert the struct Foo
from above to individual
bindings like so:
fn foo() {
let s = vec![1, 2, 3];
let p = &s;
}
Then, we could wrap both of them in a struct that only has references and use that instead:
struct AlmostFoo<'a> {
s: &'a Vec<u8>,
p: &'a Vec<u8>,
}
fn make_foo() {
let s = vec![1, 2, 3];
let p = &s;
let foo = AlmostFoo { s, p };
do_stuff(foo); // call a function that expects an AlmostFoo
}
Of course make_foo()
cannot return an AlmostFoo
instance since it would be
referencing values from its stack, but what it can do is call other functions
and pass an AlmostFoo
to them. In other words, as long as the code that wants
to use AlmostFoo
is above make_foo()
we can use this technique and work
with almost-self-references.
This is pretty restrictive though. Ideally we'd lke to be able return some owned value and be free to move it around, put it on the heap, etc.
Note: The description of async stacks bellow is not what actually happens in rustc but is enough to illustrate the point.
escher
's API does make use that the desired values are held across an await point to force them to be included in the generated Future.
As we saw, it is impossible to return an AlmostFoo
instance since it
references values from the stack. But what if we could freeze the stack after
an AlmostFoo
instance got constructed and then returned the whole stack?
Well, there is no way for a regular function to capture its own stack and
return it but that is exactly what the async/await transformation does! Let's
make make_foo
from above async and also make it never terminate:
async fn make_foo() {
let s = vec![1, 2, 3];
let p = &s;
let foo = AlmostFoo { s, p };
std::future::pending().await
}
Now when someone calls make_foo()
what they get back is some struct that
implements Future. This struct is in fact a representation of the stack of
make_foo
at its initial state, i.e in the state that the function has not be
called yet.
What we need to do now is to step the execution of the returned Future until
the instance of AlmostFoo
is constructed. In this case we know that there is
a single await point so we only need to poll the Future once. Before we do that
though we need to put it in a Pinned Box to ensure that as we poll the future
no moves will occur. This is the same restriction as with normal function but
with async it is enforced using the Pin<P>
type.
let foo = make_foo(); // construct a stack that will eventually make an AlmostFoo in it
let mut foo = Box::pin(foo_fut); // pin it so that it never moves again
foo.poll(); // poll it once
// now we know that somewhere inside `foo` there is a valid AlmostFoo instance!
We're almost there! We now have an owned value, the future, that somewhere inside it has an AlmostFoo instance. However we have no way of retrieving the exact memory location of it or accessing it in any way. The Future is opaque.
escher
builds upon the techniques described above and provides a solution for
getting the pointer from within the opaque future struct. Each Escher<T>
instance holds a Pinned Future and a raw pointer to T. The pointer to T is
computed by polling the Future just enough times for the desired T to be
constructed.
As its API, it provides the as_ref()
and as_mut()
methods that unsafely
turn the raw pointer to T into a &T with its lifetime bound to the lifetime of
Escher<T>
itself. This ensures that the future will outlive any usage of the
self-reference!
Thank you for reading this far! If you would like to learn how escher uses the above concepts in detail please take a look at the implementation.
Licensed under either of
- Apache License, Version 2.0, (LICENSE-APACHE or https://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or https://opensource.org/licenses/MIT)
at your option.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.