74 releases

0.0.124 Sep 8, 2024
0.0.123 Jul 11, 2024
0.0.122 May 17, 2024
0.0.108 Mar 30, 2024
0.0.91 Nov 13, 2023

#1161 in Encoding

Download history 2627/week @ 2024-07-29 2170/week @ 2024-08-05 3176/week @ 2024-08-12 2063/week @ 2024-08-19 2565/week @ 2024-08-26 3165/week @ 2024-09-02 2776/week @ 2024-09-09 2354/week @ 2024-09-16 3123/week @ 2024-09-23 2936/week @ 2024-09-30 3010/week @ 2024-10-07 3187/week @ 2024-10-14 2091/week @ 2024-10-21 2226/week @ 2024-10-28 1916/week @ 2024-11-04 4514/week @ 2024-11-11

10,794 downloads per month
Used in musli-tests

MIT/Apache

530KB
9K SLoC

musli-zerocopy

github crates.io docs.rs build status

Refreshingly simple, blazingly fast zero copy primitives by Müsli.

This provides a basic set of tools to deal with types which do not require copying during deserialization. You define the T, and we provide the safe &[u8] <-> &T conversions.

Reading a zero-copy structure has full #[no_std] support. Constructing ones currently requires the alloc feature to be enabled.

use musli_zerocopy::{buf, Ref, ZeroCopy};

#[derive(ZeroCopy)]
#[repr(C)]
struct Person {
    age: u8,
    name: Ref<str>,
}

let buf = buf::aligned_buf::<Person>(include_bytes!("author.bin"));
let person = Person::from_bytes(&buf[..])?;

assert_eq!(person.age, 35);
// References are incrementally validated.
assert_eq!(buf.load(person.name)?, "John-John");

For a detailed overview of how this works, see the ZeroCopy and its corresponding ZeroCopy derive. There's also a high level guide just below.

This crate also includes a couple of neat high level data structures you might be interested in:

  • phf provides maps and sets based on phf crate, or perfect hash functions.
  • swiss is a port of the hashbrown crate which is a Google SwissTable implementation.
  • trie is an implementation of a prefix-trie, which supports efficient multi-value byte-prefixed lookups.

Finally if you're interested in the performance of musli-zerocopy you should go to benchmarks. I will be extending this suite with more zero-copy types, but for now we have a clear lead in the use cases I've tested it for.

This is because:

  • Zero-copy doesn't incur a deserialization overhead if done correctly. You take bytes in one place, validate them, and treat them as the destination type. There are only so many ways this can be done;
  • Padding has been implemented and optimized in such a way that it mostly generates the equivalent code you'd write by hand, and;
  • Incremental validation means that you only need to pay for what you're accessing. So for random access we only need to validate the parts that are being accessed.

Overview:


Why should I consider musli-zerocopy over X?

Since this is the first question anyone will ask, here is how we differ from other popular libraries:

  • zerocopy doesn't support padded types[^padded], bytes to reference conversions, or conversions which does not permit decoding types unless all bit patterns can be inhabited by zeroes[^zeroes]. We still want to provide more of a complete toolkit that you'd need to build and interact with complex data structures like we get through the phf and swiss modules. This crate might indeed at some point make use of zerocopy's traits.
  • rkyv operates on #[repr(Rust)] types and from this derives an optimized Archived variation for you. This library lets you build the equivalent of the Archived variant directly and the way you interact with the data model doesn't incur the cost of validation up front. With rkyv it took my computer 100% of a CPU core and about half a second to load 12 million dictionary entries^dictionary, which is a cost that is simply not incurred by incrementally validating. Not validating is not an option since that would be wildly unsound - your application would be vulnerable to malicious dictionary files.

[^padded]: This is on zerocopy's roadmap, but it fundamentally doesn't play well with the central FromBytes / ToBytes pair of traits

[^zeroes]: FromBytes extends FromZeroes


Guide

Zero-copy in this library refers to the act of interacting with data structures that reside directly in &[u8] memory without the need to first decode them.

Conceptually it works a bit like this.

Say you want to store the string "Hello World!".

use musli_zerocopy::OwnedBuf;

let mut buf = OwnedBuf::new();
let string = buf.store_unsized("Hello World!");
let reference = buf.store(&string);

assert_eq!(reference.offset(), 12);

This would result in the following buffer:

0000: "Hello World!"
// Might get padded to ensure that the size is aligned by 4 bytes.
0012: offset -> 0000
0016: size -> 12

What we see at offset 0016 is an 8 byte Ref<str>. The first field stores the offset where to fetch the string, and the second field the length of the string.

Let's have a look at a Ref<[u32]> next:

use musli_zerocopy::{Ref, OwnedBuf};

let mut buf = OwnedBuf::new();
let slice: Ref<[u32]> = buf.store_slice(&[1, 2, 3, 4]);
let reference = buf.store(&slice);

assert_eq!(reference.offset(), 16);

This would result in the following buffer:

0000: u32 -> 1
0004: u32 -> 2
0008: u32 -> 3
0012: u32 -> 4
0016: offset -> 0000
0020: length -> 4

At address 0016 we store two fields which corresponds to a Ref<[u32]>.

Next lets investigate an example using a Custom struct:

use core::mem::size_of;
use musli_zerocopy::{OwnedBuf, Ref, ZeroCopy};

#[derive(ZeroCopy)]
#[repr(C)]
struct Custom {
    field: u32,
    string: Ref<str>,
}

let mut buf = OwnedBuf::new();

let string = buf.store_unsized("Hello World!");
let custom = buf.store(&Custom { field: 42, string });

// The buffer stores both the unsized string and the Custom element.
assert!(buf.len() >= 24);
// We assert that the produced alignment is smaller or equal to 8
// since we'll be relying on this below.
assert!(buf.requested() <= 8);

This would result in the following buffer:

0000: "Hello World!"
0012: u32 -> 42
0016: offset -> 0000
0020: size -> 12

Our struct starts at address 0012, first we have the u32 field, and immediately after that we have the string.


Reading data

Later when we want to use the type, we take the buffer we've generated and include it somewhere else.

There's a few pieces of data (lets call it DNA) we need to have to read a type back from a raw buffer:

  • The alignment of the buffer. Which you can read through the requested(). On the receiving end we need to ensure that the buffer follow this alignment. Dynamically this can be achieved by loading the buffer using aligned_buf(bytes, align). Other tricks include embedding a static buffer inside of an aligned newtype which we'll showcase below. Networked applications might simply agree to use a particular alignment up front. This alignment has to be compatible with the types being coerced.
  • The endianness of the machine which produced the buffer. Any numerical elements will in native endian ordering, so they would have to be adjusted on the read side if it differ.
  • The type definition which is being read which implements ZeroCopy. This is Custom above. The ZeroCopy derive ensures that we can safely coerce a buffer into a reference of the type. The data might at worst be garbled, but we can never do anything unsound while using safe APIs.
  • The offset at where the ZeroCopy structure is read. To read a structure we combine a pointer and a type into a Ref instance.

If the goal is to both produce and read the buffer on the same system certain assumptions can be made. And if those assumptions turn out to be wrong the worst outcome will only ever be an error as long as you're using the safe APIs or abide by the safety documentation of the unsafe APIs.

Info A note on sending data over the network. This is perfectly doable as long as you include the alignment of the buffer and the endianness of the data structure. Both of these can be retrieved:

# use musli_zerocopy::OwnedBuf;
let buf = OwnedBuf::new();

/* write something */

let is_little_endian = cfg!(target_endian = "little");
let alignment = buf.requested();

The following is an example of reading the type directly out of a newtype aligned &'static [u8] buffer:

use core::mem::size_of;
use musli_zerocopy::Buf;

// Helper to force the static buffer to be aligned like `A`.
#[repr(C)]
struct Align<A, T: ?Sized>([A; 0], T);

static BYTES: &Align<u64, [u8]> = &Align([], *include_bytes!("custom.bin"));

let buf = Buf::new(&BYTES.1);

// Construct a pointer into the buffer.
let custom = Ref::<Custom>::new(BYTES.1.len() - size_of::<Custom>());

let custom: &Custom = buf.load(custom)?;
assert_eq!(custom.field, 42);
assert_eq!(buf.load(custom.string)?, "Hello World!");

Writing data at offset zero

Most of the time you want to write data where the first element in the buffer is the element currently being written.

This is useful because it satisfies the last requirement above, the offset at where the struct can be read will then simply be zero, and all the data it depends on are stored at subsequent offsets.

use musli_zerocopy::OwnedBuf;
use musli_zerocopy::mem::MaybeUninit;

let mut buf = OwnedBuf::new();
let reference: Ref<MaybeUninit<Custom>> = buf.store_uninit::<Custom>();

let string = buf.store_unsized("Hello World!");

buf.load_uninit_mut(reference).write(&Custom { field: 42, string });

let reference = reference.assume_init();
assert_eq!(reference.offset(), 0);

Portability

By default archives will use the native ByteOrder. In order to construct and load a portable archive, the byte order in use has to be explicitly specified.

This is done by specifying the byte order in use during buffer construction and expliclty setting the E parameter in types which received it such as Ref<T, E, O>.

We can start of by defining a fully Portable archive structure, which received both size and ByteOrder. Note that it could also just explicitly specify a desired byte order but doing it like this makes it maximally flexible as an example:

use musli_zerocopy::{Size, ByteOrder, Ref, Endian, ZeroCopy};

#[derive(ZeroCopy)]
#[repr(C)]
struct Archive<E, O>
where
    E: ByteOrder,
    O: Size
{
    string: Ref<str, E, O>,
    number: Endian<u32, E>,
}

Building a buffer out of the structure is fairly straight forward, OwnedBuf has the with_byte_order::<E>() method which allows us to specify a "sticky" ByteOrder to use in types which interact with it during construction:

use musli_zerocopy::{endian, Endian, OwnedBuf};
let mut buf = OwnedBuf::new()
    .with_byte_order::<endian::Little>();

let first = buf.store(&Endian::le(42u16));
let portable = Archive {
    string: buf.store_unsized("Hello World!"),
    number: Endian::new(10),
};
let portable = buf.store(&portable);

assert_eq!(&buf[..], &[
    42, 0, // 42u16
    72, 101, 108, 108, 111, 32, 87, 111, 114, 108, 100, 33, // "Hello World!"
    0, 0, // padding
    2, 0, 0, 0, 12, 0, 0, 0, // Ref<str>
    10, 0, 0, 0 // 10u32
]);

let portable = buf.load(portable)?;

let mut buf = OwnedBuf::new()
    .with_byte_order::<endian::Big>();

let first = buf.store(&Endian::be(42u16));
let portable = Archive {
    string: buf.store_unsized("Hello World!"),
    number: Endian::new(10),
};
let portable = buf.store(&portable);

assert_eq!(&buf[..], &[
    0, 42, // 42u16
    72, 101, 108, 108, 111, 32, 87, 111, 114, 108, 100, 33, // "Hello World!"
    0, 0, // padding
    0, 0, 0, 2, 0, 0, 0, 12, // Ref<str>
    0, 0, 0, 10 // 10u32
]);

let portable = buf.load(portable)?;

Limits

Offset, the size of unsized values, and slice lengths are all limited to 32-bit. The system you're using must have a usize type which is at least 32-bits wide. This is done to save space by default.

The pointer width on the system is checked at compile time, while trying to use an offset or a size larger than 2^32 will result in a panic.

Example of using an address larger than 2^32 causing a panic:

Ref::<Custom>::new(1usize << 32);

Example panic using a Ref<\[T\]> with a length larger than 2^32:

Ref::<[Custom]>::with_metadata(0, 1usize << 32);

Example panic using an Ref<str> value with a size larger than 2^32:

Ref::<str>::with_metadata(0, 1usize << 32);

If you want to address data larger than this limit, it is recommended that you partition your dataset into 32-bit addressable chunks.

If you really want to change this limit, you can modify it by setting the default O parameter on the various Size-dependent types:

The available Size implementations are:

  • u32 for 32-bit sized pointers (the default).
  • usize for target-dependently sized pointers.
// These no longer panic:
let reference = Ref::<Custom, Native, usize>::new(1usize << 32);
let slice = Ref::<[Custom], Native, usize>::with_metadata(0, 1usize << 32);
let unsize = Ref::<str, Native, usize>::with_metadata(0, 1usize << 32);

To initialize an OwnedBuf with a custom Size, you can use OwnedBuf::with_size:

use musli_zerocopy::OwnedBuf;
use musli_zerocopy::buf::DefaultAlignment;

let mut buf = OwnedBuf::with_capacity_and_alignment::<DefaultAlignment>(0)
    .with_size::<usize>();

The Size you've specified during construction of an OwnedBuf will then carry into any pointers it return:

use musli_zerocopy::{DefaultAlignment, OwnedBuf, Ref, ZeroCopy};
use musli_zerocopy::endian::Native;

#[derive(ZeroCopy)]
#[repr(C)]
struct Custom {
    reference: Ref<u32, Native, usize>,
    slice: Ref::<[u32], Native, usize>,
    unsize: Ref::<str, Native, usize>,
}

let mut buf = OwnedBuf::with_capacity(0)
    .with_size::<usize>();

let reference = buf.store(&42u32);
let slice = buf.store_slice(&[1, 2, 3, 4]);
let unsize = buf.store_unsized("Hello World");

buf.store(&Custom { reference, slice, unsize });

Dependencies

~0.5–1MB
~21K SLoC