speice.io/_drafts/case-study-borrow-checker.md

146 lines
6.5 KiB
Markdown
Raw Normal View History

2019-05-03 16:19:32 -04:00
---
layout: post
title: "A Case Study in Borrow Checking"
description: "...and some practical lessons learned."
category:
tags: [rust]
---
I'm convinced that WebSockets are a gateway drug. The specification is reasonably easy to understand, and implementations are an opportunity to both dig into the lower-level details of networking code and experiment with new techniques. It's essentially [writing](https://www.youtube.com/watch?v=HyzD8pNlpwI) [a](https://cturt.github.io/cinoop.html) [Gameboy](https://blog.rekawek.eu/2017/02/09/coffee-gb/) [emulator](https://djhworld.github.io/post/2018/09/21/i-ported-my-gameboy-color-emulator-to-webassembly/), but for network code instead of emulation.
At least, that's how I'm approaching it. While there are [existing](https://github.com/housleyjk/ws-rs) [implementations](https://github.com/websockets-rs/rust-websocket) of the protocol for Rust, writing a WebSocket library is an opportunity for me to experiment with parser [combinators](https://github.com/Geal/nom) and [generators](https://github.com/kaitai-io/kaitai_struct), and maybe have something to show at the end of it. Recently, I've been adding support for Rust to the [Kaitai Struct](http://kaitai.io/) project so that I can generate the parser from a schema, rather than writing one by hand. But before we can generate a parser using Kaitai, we need a runtime library. This is a typical pattern in code generation; the generated code relies on a "standard library" of functionality similar to how programming languages have their own standard library.
What makes this parser runtime difficult to implement in Rust is the performance concerns; we don't want to allocate new `Vec<u8>` buffers and copy data around when it's not necessary. Especially in networking code, these types of "zero-copy" operations are critical to performance. And because we're not interested in modifying the data stream, references make a lot of sense! However, that means there's a good potential to hit issues with the borrow checker; making sure all the structures being parsed use the stream correctly is difficult. As a result, I hit a lot of issues with the borrow checker, and wanted to detail what I learned about not only how to *avoid* fighting the borrow checker, but how to *work with* the borrow checker.
# Design Inspiration - C++
So how exactly does one go about building such a runtime? In this case, we'll start by looking at Kaitai's [C++ support](https://github.com/kaitai-io/kaitai_struct_cpp_stl_runtime) for inspiration, and see if we can adapt that to Rust. There's even an [ownership guide](http://doc.kaitai.io/lang_cpp_stl.html#_ownership_model) detailing the rules for how the C++ runtime thinks about ownership!
This article will use a toy schema for illustrating lifetimes:
```yaml
meta:
id: toy
title: Toy Schema
endian: be
seq:
- id: slice_size
type: u1
- id: child_structure
type: child
types:
child:
seq:
- id: slice
size: _parent.slice_size
- id: grandchild_structure
type: grandchild
grandchild:
seq:
- id: slice
size: _root.slice_size
```
The parser will operate like this:
1. Read a single byte (`u8` in Rust) from a stream, and store that in `slice_size`
2. Read a child structure. First, read a byte slice whose size is the parent structure's `slice_size`, then read the grandchild
3. Read a granchild structure by taking a byte slice whose size is the root structure's `slice_size`
So let's start by generating the C++ code corresponding to our specification (edited for clarity):
**toy.h**
```cpp
class toy_t : public kaitai::kstruct {
public:
class child_t;
class grandchild_t;
toy_t(kaitai::kstream* p__io, kaitai::kstruct* p__parent = nullptr, toy_t* p__root = nullptr);
~toy_t();
class child_t : public kaitai::kstruct {
public:
child_t(kaitai::kstream* p__io, toy_t* p__parent = nullptr, toy_t* p__root = nullptr);
~child_t();
private:
std::string m_slice;
std::unique_ptr<grandchild_t> m_grandchild_structure;
toy_t* m__root;
toy_t* m__parent;
};
class grandchild_t : public kaitai::kstruct {
public:
grandchild_t(kaitai::kstream* p__io, toy_t::child_t* p__parent = nullptr, toy_t* p__root = nullptr);
~grandchild_t();
private:
std::string m_slice;
toy_t* m__root;
toy_t::child_t* m__parent;
};
private:
uint8_t m_initial_byte;
std::unique_ptr<child_t> m_child_structure;
toy_t* m__root;
kaitai::kstruct* m__parent;
};
```
**toy.cpp**
```cpp
#include "toy.h"
toy_t::toy_t(kaitai::kstream* p__io, kaitai::kstruct* p__parent, toy_t* p__root) : kaitai::kstruct(p__io) {
m__parent = p__parent;
m__root = this;
m_child_structure = nullptr;
_read();
}
void toy_t::_read() {
m_slice_size = m__io->read_u1();
m_child_structure = std::unique_ptr<child_t>(new child_t(m__io, this, m__root));
}
toy_t::child_t::child_t(kaitai::kstream* p__io, toy_t* p__parent, toy_t* p__root) : kaitai::kstruct(p__io) {
m__parent = p__parent;
m__root = p__root;
m_grandchild_structure = nullptr;
_read();
}
void toy_t::child_t::_read() {
m_slice = m__io->read_bytes(_parent()->slice_size());
m_grandchild_structure = std::unique_ptr<grandchild_t>(new grandchild_t(m__io, this, m__root));
}
toy_t::grandchild_t::grandchild_t(kaitai::kstream* p__io, toy_t::child_t* p__parent, toy_t* p__root) : kaitai::kstruct(p__io) {
m__parent = p__parent;
m__root = p__root;
_read();
}
void toy_t::grandchild_t::_read() {
m_slice = m__io->read_bytes(_root()->slice_size());
}
```
Now, let's think about ownership as we look at the code:
- Each parent structure (`toy_t` and `child_t`) expresses ownership of its children through `std::unique_ptr<>`.
- Because children can refer to parents through `m__parent` and `m__root`, we have a reference cycle that will be difficult to express in Rust.
- Everyone stores a reference to `kaitai::kstream`, but nobody owns it.
- Structures own their data using `std::string` ([`read_bytes` implementation](https://github.com/kaitai-io/kaitai_struct_cpp_stl_runtime/blob/1ea056ad053b438e1609fe84e71b1d306777492d/kaitai/kaitaistream.cpp#L347-L361));
this prevents issues if the stream (`m__io`) gets destroyed, but also introduces an extra allocation and copy that Rust can avoid if we convince the borrow checker that structures won't outlive the stream.
- The root structure (`toy_t`) stores a reference to itself; it's thus unsafe to copy or move.
With all that in mind, let's talk about ownership in the Rust runtime.