"I'm at the buffer pool. I'm at the MPSC channel. I'm at the combination MPSC channel and buffer pool."
thingbuf
is a lock-free array-based concurrent ring buffer that allows access
to slots in the buffer by reference. It's also asynchronous
and blocking bounded MPSC channels implemented using
the ring buffer.
If you want a high-throughput bounded MPSC channel that allocates only on
channel creation. Some MPSC channels have good throughput. Some other MPSC
channels won't allocate memory per-waiter. thingbuf::mpsc
has both.
thingbuf::mpsc
is a competitive choice for a general-purpose
MPSC channel in most use cases.
Both asynchronous and blocking
MPSC channels are available, so thingbuf
can be used in place
of asynchronous channels like futures::channel::mpsc
and blocking
channels like std::sync::mpsc::sync_channel
.
If you can't allocate or you need to build with #![no_std]
because
you're working on embedded systems or other bare-metal software. Thingbuf
provides a statically-allocated MPSC channel and a
statically-allocated lock-free queue. These can be placed in a
static
initializer and used without requiring any runtime allocations.
You want to use the same MPSC channel with and without std
. Thingbuf's
asynchronous MPSC channel provides an identical API and feature set regardless
of whether or not the "std" feature flag is enabled. If you're writing a library that
needs to conditionally support #![no_std]
, and you need an asynchronous MPSC
channel, it might be easier to use thingbuf::mpsc
in both cases, rather
than switching between separate std
and #![no_std]
channel
implementations.
It's equally important to discuss when thingbuf
should not be used. Here are
some cases where you might be better off considering other options:
You need a really, really, absurdly high bound and you're not going to be
near it most of the time. If you want to set a very, very high bound on a
bounded MPSC channel, and the channel will typically never be anywhere near
that full, thingbuf::mpsc
might not be the best choice.
Thingbuf's channels will allocate an array with length equal to the capacity
as soon as they're constructed. This improves performance by avoiding
additional allocations, but if you need to set very high bounds, you might
prefer a channel implementation that only allocates memory for messages as
it's needed (such as tokio::sync::mpsc
).
You need a blocking channel with a select
operation.
I'm probably not going to implement it. I may accept a PR if you raise it.
If you need a synchronous channel with this kind of functionality,
crossbeam-channel
is probably a good choice.
You want an unbounded channel. I'm not going to write an unbounded channel. Unbounded channels are evil.
This crate's API and documentation makes a distinction between the terms "queue" and "channel". The term queue will refer to the queue abstract data type in general — any first-in, first-out data structure is a queue.
The term channel will refer to a subtype of concurrent queue that also functions as a synchronization primitive. A channel is a queue which can be shared between multiple threads or asynchronous tasks, and which allows those threads or tasks to wait for elements to be added or removed from the queue.
In the Rust standard library, the std::collections::VecDeque
type
is an example of a queue that is not a channel: it is a first-in, first-out data
structure, but it cannot be concurrently enqueued to and dequeued from by
multiple threads or tasks. In comparison, the types in the std::sync::mpsc
module provide a prototypical example of channels, as they serve as
synchronization primitives for cross-thread communication.
To get started using thingbuf
, add the following to your Cargo.toml
:
[dependencies]
thingbuf = "0.1"
By default, thingbuf
depends on the Rust standard library, in order to
implement APIs such as synchronous (blocking) channels. In #![no_std]
projects, the std
feature flag must be disabled:
[dependencies]
thingbuf = { version = "0.1", default-features = false }
With the std
feature disabled, thingbuf
will depend only on libcore
. This
means that APIs that require dynamic memory allocation will not be enabled.
Statically allocated channels and queues are
available for code without a memory allocator, if the static
feature flag is
enabled:
[dependencies]
thingbuf = { version = "0.1", default-features = false, features = ["static"] }
However, if a memory allocator is available, #![no_std]
code can also enable
the alloc
feature flag to depend on liballoc
:
[dependencies]
thingbuf = { version = "0.1", default-features = false, features = ["alloc"] }
liballoc
(but not libstd
). This
enables thingbuf
queues and asynchronous channels where the size of the
channel is determined at runtime.thingbuf
queues and channels. These can be used
without dynamic memory allocation when the size of a queue or channel is known
at compile-time.thingbuf
is built against the latest stable release. The minimum supported
version is Rust 1.57. The current thingbuf
version is not guaranteed to build on Rust
versions earlier than the minimum supported version.
Some feature flags may require newer Rust releases. For example, the "static" feature flag requries Rust 1.60+.
Q: Why did you make this?
A: For tracing
, I wanted to be able to send formatted log lines to a
dedicated worker thread that writes them to a file. Right now, we do this
using crossbeam-channel
. However, this has the sad disadvantage that we have
to allocate String
s, send them through the channel to the writer, and
immediately drop them. It would be nice to do this while reusing those
allocations. Thus...StringBuf
.
Q: Is it lock-free?
A: Extremely.
Q: Why is there only a bounded variant?
A: Because unbounded queues are of the Devil.
Q: Isn't this just a giant memory leak?
A: If you use it wrong, yes.
Q: Why is it called that?
A: Originally, I imagined it as a kind of ring buffer, so (as a pun on "ringbuf"), I called it "stringbuf". Then, I realized you could do this with more than just strings. In fact, it can be generalized to arbitrary...things. So, "thingbuf".