3 releases
0.3.4 | Jul 27, 2024 |
---|---|
0.3.3 | Nov 13, 2023 |
0.3.2 | Nov 18, 2021 |
0.3.1 |
|
#91 in Data structures
5,770 downloads per month
Used in 16 crates
(4 directly)
235KB
5K
SLoC
Slice Ring Buffer
A double-ended queue that
Deref
s into a slice, also known as a ring buffer or circular buffer.
⚠️ Notice ⚠️
This is a fork of SliceDequeue that adds a security patch due to the original project no longer being maintained. I will keep the repo updated with security patches etc. But no active feature development will be done. Pull requests are welcome
Advantages
The main advantages of SliceRingBuffer
are:
-
nicer API: since it
Deref
s to a slice, all operations that work on slices (likesort
) "just work" forSliceRingBuffer
. -
efficient: as efficient as a slice (iteration, sorting, etc.), more efficient in general than
VecDeque
.
Platform Support
Windows, Linux, MacOS and every other unix-like OS is supported (although maybe untested). The following targets are known to work and pass all tests:
Linux
- aarch64-unknown-linux-gnu
- arm-unknown-linux-gnueabi
- arm-unknown-linux-musleabi
- armv7-unknown-linux-gnueabihf
- armv7-unknown-linux-musleabihf
- i586-unknown-linux-gnu
- i686-unknown-linux-gnu
- i686-unknown-linux-musl
- mips-unknown-linux-gnu
- mips64-unknown-linux-gnuabi64
- mips64el-unknown-linux-gnuabi64
- mipsel-unknown-linux-gnu
- powerpc-unknown-linux-gnu
- powerpc64-unknown-linux-gnu
- powerpc64le-unknown-linux-gnu
- x86_64-unknown-linux-gnu
- x86_64-unknown-linux-musl
- aarch64-linux-android
- arm-linux-androideabi
- armv7-linux-androideabi
- x86_64-linux-android
MacOS X
- i686-apple-darwin
- x86_64-apple-darwin
Windows
- x86_64-pc-windows-msvc
Drawbacks
The main drawbacks of SliceRingBuffer
are:
-
"constrained" platform support: the operating system must support virtual memory. In general, if you can use
std
, you can useSliceRingBuffer
. -
global allocator bypass:
SliceRingBuffer
bypasses Rust's global allocator / it is its own memory allocator, talking directly to the OS. That is, allocating and growingSliceRingBuffer
s always involve system calls, while aVecDeque
backed-up by a global allocator might receive memory owned by the allocator without any system calls at all. -
smallest capacity constrained by the allocation granularity of the OS: some operating systems allow
SliceRingBuffer
to allocate memory in 4/8/64 kB chunks.
When shouldn't you use it? In my opinion, if
- you need to target
#[no_std]
, or - you can't use it (because your platform doesn't support it)
you must use something else. If.
- your ring-buffer's are very small,
then by using SliceRingBuffer
you might be trading memory for performance. Also,
- your application has many short-lived ring-buffers,
the cost of the system calls required to set up and grow the SliceRingBuffer
s
might not be amortized by your application (update: there is a pull-request open
that caches allocations in thread-local heaps when the feature use_std
is
enabled significantly improving the performance of short-lived ring-buffers, but
it has not been merged yet). Whether any of these trade-offs are worth it or not
is application dependent, so don't take my word for it: measure.
How it works
The double-ended queue in the standard library (VecDeque
) is implemented
using a growable ring buffer (0
represents uninitialized memory, and T
represents one element in the queue):
// [ 0 | 0 | 0 | T | T | T | 0 ]
// ^:head ^:tail
When the queue grows beyond the end of the allocated buffer, its tail wraps around:
// [ T | T | 0 | T | T | T | T ]
// ^:tail ^:head
As a consequence, VecDeque
cannot Deref
into a slice, since its elements
do not, in general, occupy a contiguous memory region. This complicates the
implementation and its interface (for example, there is no as_slice
method -
the as_slices
method returns a pair of slices) and has negative performance
consequences (e.g. need to account for wrap around while iterating over the
elements).
This crates provides SliceRingBuffer
, a double-ended queue implemented with
a growable virtual ring-buffer.
A virtual ring-buffer implementation is very similar to the one used in
VecDeque
. The main difference is that a virtual ring-buffer maps two
adjacent regions of virtual memory to the same region of physical memory:
// Virtual memory:
//
// __________region_0_________ __________region_1_________
// [ 0 | 0 | 0 | T | T | T | 0 | 0 | 0 | 0 | T | T | T | 0 ]
// ^:head ^:tail
//
// Physical memory:
//
// [ 0 | 0 | 0 | T | T | T | 0 ]
// ^:head ^:tail
That is, both the virtual memory regions 0
and 1
above (top) map to the same
physical memory (bottom). Just like VecDeque
, when the queue grows beyond the
end of the allocated physical memory region, the queue wraps around, and new
elements continue to be appended at the beginning of the queue. However, because
SliceRingBuffer
maps the physical memory to two adjacent memory regions, in virtual
memory space the queue maintais the ilusion of a contiguous memory layout:
// Virtual memory:
//
// __________region_0_________ __________region_1_________
// [ T | T | 0 | T | T | T | T | T | T | 0 | T | T | T | T ]
// ^:head ^:tail
//
// Physical memory:
//
// [ T | T | 0 | T | T | T | T ]
// ^:tail ^:head
Since processes in many Operating Systems only deal with virtual memory
addresses, leaving the mapping to physical memory to the CPU Memory Management
Unit (MMU), SliceRingBuffer
is able to Deref
s into a slice in those systems.
This simplifies SliceRingBuffer
's API and implementation, giving it a performance
advantage over VecDeque
in some situations.
In general, you can think of SliceRingBuffer
as a Vec
with O(1)
pop_front
and amortized O(1)
push_front
methods.
License
This project is licensed under either of
- Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
at your option.
Contribution
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in SliceRingBuffer by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.