Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions src/SUMMARY.md
Original file line number Diff line number Diff line change
Expand Up @@ -445,6 +445,11 @@
- [Serializer: implement Struct](idiomatic/leveraging-the-type-system/typestate-pattern/typestate-generics/struct.md)
- [Serializer: implement Property](idiomatic/leveraging-the-type-system/typestate-pattern/typestate-generics/property.md)
- [Serializer: Complete implementation](idiomatic/leveraging-the-type-system/typestate-pattern/typestate-generics/complete.md)
- [Borrow checking invariants](idiomatic/leveraging-the-type-system/borrow-checker-invariants.md)
- [Generalizing "Ownership"](idiomatic/leveraging-the-type-system/borrow-checker-invariants/generalizing-ownership.md)
- [Single-use values](idiomatic/leveraging-the-type-system/borrow-checker-invariants/single-use-values.md)
- [Aliasing XOR Mutability](idiomatic/leveraging-the-type-system/borrow-checker-invariants/aliasing-xor-mutability.md)
- [Lifetime Relationships & External Resources](idiomatic/leveraging-the-type-system/borrow-checker-invariants/lifetime-connections.md)

---

Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,80 @@
---
minutes: 10
---

# Using the Borrow checker to enforce Invariants

The logic of the borrow checker, while tied to "memory ownership", can be
abstracted away from this central use case to model other problems and prevent
API misuse.

```rust,editable
// Doors can be open or closed, and you need the right key to lock or unlock
// one. Modelled with a Shared key and Owned door.
pub struct DoorKey {
pub key_shape: u32,
}
pub struct LockedDoor {
lock_shape: u32,
}
pub struct OpenDoor {
lock_shape: u32,
}

fn open_door(key: &DoorKey, door: LockedDoor) -> Result<OpenDoor, LockedDoor> {
if door.lock_shape == key.key_shape {
Ok(OpenDoor { lock_shape: door.lock_shape })
} else {
Err(door)
}
}

fn close_door(key: &DoorKey, door: OpenDoor) -> Result<LockedDoor, OpenDoor> {
if door.lock_shape == key.key_shape {
Ok(LockedDoor { lock_shape: door.lock_shape })
} else {
Err(door)
}
}

fn main() {
let key = DoorKey { key_shape: 7 };
let closed_door = LockedDoor { lock_shape: 7 };
let opened_door = open_door(&key, closed_door);
if let Ok(opened_door) = opened_door {
println!("Opened the door with key shape '{}'", key.key_shape);
} else {
eprintln!(
"Door wasn't opened! Your key only opens locks with shape '{}'",
key.key_shape
);
}
}
```

<details>

<!-- TODO: link to typestate when that gets merged. -->

- The borrow checker has been used to prevent use-after-free and multiple
mutable references up until this point, and we've used types to shape and
restrict use of APIs already using the "typestate" pattern.

- This example uses the ownership & borrowing rules to model the locking and
unlocking of a door. We can try to open a door with a key, but if it's the
wrong key the door is still closed (here represented as an error) and the key
persists regardless.

- The rules of the borrow checker exist to prevent developers from accessing,
changing, and holding onto data in memory in unpredictable ways without being
so restrictive that it would prevent _writing software_. The underlying
logical system does not "know" what memory is. All it does is enforce a
specific set of rules of how different operations affect what later operations
are possible.

- Those rules can apply to many other cases: We can piggy-back onto the rules of
the borrow checker to design APIs to be harder or impossible to misuse, even
when there's little or no "memory safety" concerns in the problem domain. This
section will walk through some of those different domains.

</details>
Original file line number Diff line number Diff line change
@@ -0,0 +1,99 @@
---
minutes: 15
---

# Mutually Exclusive References, or "Aliasing XOR Mutability"

We can use the mutual exclusion of `&T` and `&mut T` references for a single
value to model some constraints.

```rust,editable,compile_fail
pub struct Transaction(/* specifics omitted */);
pub struct QueryResult(String);

pub struct DatabaseConnection {
transaction: Transaction,
query_results: Vec<QueryResult>,
}

impl DatabaseConnection {
pub fn new() -> Self {
Self {
transaction: Transaction(/* again, specifics omitted */),
query_results: vec![],
}
}
pub fn get_transaction(&mut self) -> &mut Transaction {
&mut self.transaction
}
pub fn results(&self) -> &[QueryResult] {
&self.query_results
}
pub fn commit(&mut self) {
/* Work omitted, including sending/clearing the transaction */
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
/* Work omitted, including sending/clearing the transaction */
// Work omitted, including sending/clearing the transaction

println!("Transaction committed!")
}
}

pub fn do_something_with_transaction(transaction: &mut Transaction) {}

fn main() {
let mut db = DatabaseConnection::new();
let mut transaction = db.get_transaction();
do_something_with_transaction(transaction);
let assumed_the_transactions_happened_immediately = db.results(); // ❌🔨
do_something_with_transaction(transaction);
// Works, as the lifetime of "transaction" as a reference ended above.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The thing that ended was not the lifetime of the transaction (the transaction is still alive, it is inside the database). What ended is the borrow.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I might be misunderstanding your intent behind this example, but it does not feel realistic realistic because the database didn't get a notification that the borrow has ended, and thus could not have started executing it. In other words, it is not clear what the DatabaseConnection type achieved by being locked out while the transaction was borrowed out.

Maybe you can fix it by wrapping the reference in a custom wrapper type which implements Drop? Ideally, that drop() would also print a message (like "executing query...") to make it clear to the audience that the database can indeed fill in the result list.

let assumed_the_transactions_happened_immediately_again = db.results();
db.commit();
}
```

<details>

- Aliasing XOR Mutability means "we can have multiple immutable references, a
single mutable reference, but not both."

- This example shows how we can use the mutual exclusion of these kinds of
references to dissuade a user from reading query results while using a
transaction API.

This might happen if the user is working under the false assumption that the
queries being written to the transaction happen "immediately" rather than
being queued up and performed together.

- By borrowing one field of a struct via a method that returns a mutable /
exclusive reference we prevent access to the other fields of that struct under
a shared / non-exclusive reference until the lifetime of that borrow ends.

Note: This has to be via a method, as the compiler can reason about borrowing
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
Note: This has to be via a method, as the compiler can reason about borrowing
- The `transaction` field must be borrowed via a method, as the compiler can reason about borrowing

This is a major part (esp. since now there's a demonstration), so it deserves a top-level bullet point.

different fields in mutable/shared ways simultaneously if that borrowing is
done manually.

Demonstrate:

- Change the instances of `db.get_transaction()` and `db.results()` to manual
borrows (`&mut db.transaction` and `&db.query_results` respectively) to show
the difference in what the borrow checker allows.

- Put the non-`main` part of this example in a module to reiterate that this
manual access is not possible across module boundaries.

- As laid out in [generalizing ownership](generalizing-ownership.md) we can look
at the ways Mutable References and Shareable References interact to see if
they fit with the invariants we want to uphold for an API.

- In this case, having the query results not public and placed behind a getter
function, we can enforce the invariant "users of this API are not looking at
the query results at the same time as they are writing to a transaction."

- The "don't look at query results while building a transaction" invariant can
still be circumvented, how so?

- The user could access the transaction solely through `db.get_transaction()`,
leaving the lifetime too temporary to prevent access to `db.results()`.

- How could we avoid this by working in other concepts from "Leveraging the
Type System"?

</details>
Original file line number Diff line number Diff line change
@@ -0,0 +1,88 @@
---
minutes: 15
---

# Generalizing Ownership

The logic of the borrow checker, while modelled off "memory ownership", can be
abstracted away from that use case to model other problems where we want to
prevent API misuse.

```rust,editable,compile_fail
// An internal data type to have something to hold onto.
pub struct Internal;
// The "outer" data.
pub struct Data(Internal);

fn shared_use(value: &Data) -> &Internal {
&value.0
}
fn exclusive_use(value: &mut Data) -> &mut Internal {
&mut value.0
}
fn deny_future_use(value: Data) {}

fn main() {
let mut value = Data(Internal);
let deny_mut = shared_use(&value);
let try_to_deny_immutable = exclusive_use(&mut value); // ❌🔨
let more_mut_denial = &deny_mut;
deny_future_use(value);
let even_more_mut_denial = shared_use(&value); // ❌🔨
}
```

<details>

- This example re-frames the borrow checker rules away from references and
towards semantic meaning in non-memory-safety settings. Nothing is being
mutated, nothing is being sent across threads.

- When a new feature is introduced to users, it is often done so with a specific
idea of what it will be used for.
Comment on lines +41 to +42
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
- When a new feature is introduced to users, it is often done so with a specific
idea of what it will be used for.
- Language features are often introduced for a specific purpose.


Over time, users may develop ways of using that feature in ways that may have
not been foreseen.

In 2004, Java 5 introduced Generics with the
[main stated purpose of enabling type safe collections](https://jcp.org/en/jsr/detail?id=14).
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
[main stated purpose of enabling type safe collections](https://jcp.org/en/jsr/detail?id=14).
[main stated purpose of enabling type-safe collections](https://jcp.org/en/jsr/detail?id=14).


Since then, users and developers of the language expanded the use of generics
to other areas of type safe API design.
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
to other areas of type safe API design.
to other areas of type-safe API design.

<!-- TODO: Reference how this was adopted -->

What we aim to do here is similar: The borrow checker, after being introduced
to people with the purpose of avoiding use-after-free and data races, is being
used to model things that have nothing to do with preventing those classes of
misuse.
Comment on lines +54 to +57
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
What we aim to do here is similar: The borrow checker, after being introduced
to people with the purpose of avoiding use-after-free and data races, is being
used to model things that have nothing to do with preventing those classes of
misuse.
What we aim to do here is similar: Even though the borrow checker was introduced
to prevent use-after-free and data races, it is just another API design tool. It can be
used to model program properties that have nothing to do with preventing memory safety bugs.


- To use the borrow checker as a problem solving tool, we will need to "forget"
that the original purpose of it is to prevent mutable aliasing in the context
of preventing use-after-frees and data races, instead imagining and working
within situations where the rules are the same but the meaning is slightly
different.

- In rust's borrow checker we have access to three different ways of "taking" a
value:

<!-- TODO: actually link to the RAII section when it has been merged. -->
- Owned value `T`. Very permissive case, to the point where mutability can be
re-set, but demands that nothing else is using it in any context and drops
the value when scope ends (unless that scope returns this value) (see:
RAII.)

- Mutable Reference `&mut T`. While holding onto a mutable reference we can
still "dispatch" to methods and functions that take an immutable, shared
reference of the value but only as long as we're not aliasing immutable,
shared references to related data "after" that dispatch.

- Shared Reference `&T`. Allows aliasing but prevents mutable access while any
of these exist. We can't "dispatch" to methods and functions that take
mutable references when all we have is a shared reference.

- Remember that every `&T` and `&mut T` has a lifetime, just one the user
doesn't have to annotate or think about most of the time. We get to avoid
annotating a lot of lifetimes because the rust compiler can elide the majority
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"the rust compiler can elide"

The user can elide, the Rust compiler can expand them back.

of them. See: [Lifetime Elision](../../../lifetimes/lifetime-elision.md).

</details>
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This slide might be a place where an analogy might help some listeners understand what we are trying to get at when we say that the borrow checker and lifetimes are just another API design tool. Here's one possible analogy. WDYT?

Generics in Java were added primarily to support type-safe collections. In fact, Java 5 added generic type arguments to existing standard library collection types that were previously non-generic! So the language designers had a clear primary use case in mind. However, generics turned out to be useful in many other API designs. So it would be too narrow-minded to present Java generics as "a language feature for type-safe collections."

Similarly, the lifetimes and the borrow checker were introduced in Rust for compile-time memory safety guarantees, but their applicability in API design is broader. We (the Rust community) are still discovering design patterns and trying to understand what these tools can do for API design beyond memory safety.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Seems like a good thing to bring up, I'll pull together some references to drop in. If you've got suggestions on pieces covering this I'd be happy to hear about them, but I understand linkrot and the ephemeral nature of back-channel discussion of the time may have gotten to most of it.

Copy link
Collaborator

@gribozavr gribozavr Oct 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

JSR 14 which introduced generics lists only one goal specific to a use case, and it is "Good collections support. The core Collections APIs and similar APIs are perhaps the most important customers of genericity, so it is essential that they work well as, and with, generic classes." Furthermore, this is the #1 goal of the proposal overall.

An empirical research article that I could find, Java generics adoption: how new features are introduced, championed, or ignored studies how generics were adopted in practice. It includes a data-driven argument that the most common parameterized types are collections (the only non-collection-related type in Table 1 is Class<?>). This aligns with my intuition: the primary use case is collections, but there are other cases where generics turned out to be useful (for example, Class<?> in the standard library, TypeToken<?> in Google's Guava to work around type erasure in Java's generics, or the "recursive generics" pattern similar to CRTP in C++).

One source that I had high hopes for, ACM's History of programming languages journal, unfortunately does have a piece on Java.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wonderful, thank you!

Original file line number Diff line number Diff line change
@@ -0,0 +1,103 @@
---
minutes: 20
---

# Lifetime "Connections" & External Resources

Using `PhantomData` in conjunction with lifetimes lets us say "this value may
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This made me think, do we have a slide on PhantomData, just about the concept by itself? Should we have one?

It seems like there are a few things to say to motivate PhantomData's existence (that is, before we even introduce it on the slide), in particular that the compiler infers certain properties of a user-defined type based on its fields - this would be surprising to many people, I think.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We do not. Prior to starting work on this chapter there wasn't a mention of PhantomData in the repository at all.

If there's a place for it to exist, it's likely in a revision on the typestate work. I'll pull together something short that can slot in just before the lifetime-connections slide for now under the assumption it can be moved about as needed.

own its data, but it can only live as long as the value that generated it" in
rust's type system.

```rust,editable,compile_fail
use std::marker::PhantomData;
pub struct Tag;
pub struct ErasedData<'a> {
data: String,
_phantom: PhantomData<&'a ()>,
}
impl<'a> ErasedData<'a> {
pub fn get(&self) -> &str {
&self.data
}
}
pub struct TaggedData<T> {
data: String,
_phantom: PhantomData<T>,
}
impl<T> TaggedData<T> {
pub fn new(data: String) -> Self {
Self { data, _phantom: PhantomData }
}
pub fn consume(self) {}
pub fn get_erased(&self) -> ErasedData<'_> {
// has an owned String, but _phantom holds onto the lifetime of the
// TaggedData that created it.
ErasedData { data: self.data.clone(), _phantom: PhantomData }
}
}

fn main() {
let tagged_data: TaggedData<Tag> = TaggedData::new("Real Data".to_owned());
// Get the erased-but-still-linked data.
let erased_owned_and_linked = tagged_data.get_erased();
tagged_data.consume();
// Owned by `erased_owned_and_linked` but still connected to `tagged_data`.
println!("{}", erased_owned_and_linked.get()); // ❌🔨
}
```

<details>

- `PhantomData` lets developers "tag" types with type and lifetime parameters
that are not "really" present in the struct or enum.

`PhantomData` can be used with the Typestate pattern to have data with the
same structure i.e. `TaggedData<Start>` can have methods or trait
implementations that `TaggedData<End>` doesn't.

It can also be used to encode a connection between the lifetime of one value
and another, while both values still maintain separate owned data within them.

- This is really useful for modelling a bunch of relationships between data,
where we want to establish that while a type has owned values within it is
still connected to another piece of data and can only live as long as it.

Consider a case where you want to return owned data from a method, but you
don't want that data to live longer than the value that created it.

- Lifetimes need to come from somewhere! We can't build functions of the form
`fn lifetime_shenanigans<'a>(owned: OwnedData) -> &'b Data` (without tying
`'b` to `'a` in some way).

Lifetime elision hides where a lot of lifetimes come from, but that doesn't
mean the explicitly named lifetimes "come from nowhere."

Suggestion: Show off un-eliding the lifetimes in `get_erased` in this example.

- [`BorrowedFd`](https://rust-lang.github.io/rfcs/3128-io-safety.html#ownedfd-and-borrowedfdfd)
uses these captured lifetimes to enforce the invariant that "if this file
descriptor exists, the OS file descriptor is still open" because a
`BorrowedFd`'s lifetime parameter demands that there exists another value in
your program that has the same lifetime as it, and this has been encoded by
the API designer to mean _that value is what keeps the access to the file
open_.

Its counterpart `OwnedFd` is instead a file descriptor that closes that file
on drop.

## More to Explore

- This way of encoding information in types is _exceptionally powerful_ when
combined with unsafe, as the ways one can manipulate lifetimes becomes almost
arbitrary. This is also dangerous, but when combined with tools like external,
mechanically-verified proofs _we can safely encode cyclic/self-referential
types while encoding lifetime & safety expectations in the relevant data
types._

- The [GhostCell (2021)](https://plv.mpi-sws.org/rustbelt/ghostcell/) paper and
its [relevant implementation](https://gitlab.mpi-sws.org/FP/ghostcell) show
this kind of work off. While the borrow checker is restrictive, there are
still ways to use escape hatches and then _show that the ways you used those
escape hatches are consistent and safe._

</details>
Loading