A C POSIX multi-thread based multiple producers, single consumer (MPSC) channel library.
lib<mpsc>
is an attempt at a single process, multi-thread based multiple producers, single consumer (MPSC) channel implementation in C. It is based on POSIX's pthread facilities, making use of types such as pthread_t
, pthread_mutex_t
, and pthread_cond_t
to ensure proper and efficient message synchronization between a producer and the consumer.
I created this library as a learning activity, while trying to see if I could implement something similar to Rust's std::sync::mpsc module in C. The result turned out to be something that is not directly comparable to the Rust Standard Library's mpsc
module, but nonetheless turned out to be an insightful as well as enjoyable exercise!
Currently, this library only allows for synchronous message passing (i.e., sending), although that could change in a future version (read more).
#include <assert.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include "mpsc.h"
#define IGNORE_UNUSED(m) ((void)(m))
#define TEXT_BUFFER_SIZE (100)
#define N_PRODUCERS (8)
static void my_consumer_callback(mpsc_consumer_t *consumer, void *data, size_t n, bool closed);
static void my_producer_thread_callback(mpsc_producer_t *producer);
struct my_message
{
char text[TEXT_BUFFER_SIZE];
};
struct my_producer_thread_callback_context
{
size_t id;
};
int main(void)
{
mpsc_t *mpsc = mpsc_create((mpsc_create_params_t){
.buffer_size = sizeof(struct my_message),
.n_max_producers = N_PRODUCERS,
.consumer_callback = my_consumer_callback,
.consumer_error_callback = NULL,
.error_handling_enabled = false,
.create_and_join_thread_safety_disabled = false,
});
struct my_producer_thread_callback_context contexts[N_PRODUCERS];
for (size_t i = 0; i < N_PRODUCERS; i++)
{
contexts[i].id = i + 1;
assert(mpsc_register_producer(mpsc, my_producer_thread_callback, &contexts[i]) == MPSC_REGISTER_PRODUCER_ERROR_NONE);
}
mpsc_join(mpsc);
exit(EXIT_SUCCESS);
}
static void my_consumer_callback(mpsc_consumer_t *consumer, void *data, size_t n, bool closed)
{
IGNORE_UNUSED(consumer);
static size_t counter = 0;
counter += 1;
if (closed)
{
fprintf(stdout, "[consumer:%zu] closed\n", counter);
return;
}
if (n != sizeof(struct my_message))
{
fprintf(stderr, "[consumer] Error: Unexpected message size\n");
exit(EXIT_FAILURE);
}
struct my_message *message = data;
fprintf(stdout, "[consumer:%zu] %s\n", counter, message->text);
free(data);
}
static void my_producer_thread_callback(mpsc_producer_t *producer)
{
struct my_producer_thread_callback_context *ctx = mpsc_producer_context(producer);
struct my_message message;
sprintf(message.text, "Hello from producer #%zu!", ctx->id);
assert(mpsc_producer_send(producer, &message, sizeof(struct my_message)));
}
NOTE: There are other examples in the examples directory.
At a high level, a mpsc_t
object (i.e., a channel) is instantiated using the mpsc_create
function, which accepts a structure of type mpsc_create_params_t
as argument, which contains the configurations for the new channel. The next step is to register producers, before finally "joining" (using mpsc_join
) the channel. Note that, for most applications, the channel should be created and joined on the same thread (e.g., the main application thread).
Internally, a new thread is created for each registered producer, and an additional thread is created for the consumer. The consumer thread uses a mutex
and a condition variable
to wait on a message, which message is signaled by a producer by updating the condition variable's state (and possibly first updating the internal data buffer with data, if the message has data). The internal data buffer is protected by a mutex, and contains a memory independent copy of the message sent by the producer (in other words, a message sent by a producer, using mpsc_producer_send
, has its data copied to the internal buffer).
When the internal consumer thread gets notified of a new message, it creates a copy of the message and passes that copy as argument to the consumer callback, which callback has been implemented by the application (note: that means that the consumer callback gets executed on the internal consumer thread; not the thread calling mpsc_join
). But before calling the application defined consumer callback, the consumer thread first resets the condition variable and releases the mutex, so that other producers can send new messages.
A channel can be closed in two ways. First, the consumer callback function can itself request that the channel be closed by calling mpsc_consumer_close
on the mpsc_consumer_t
object it receives as argument. Alternatively, provided that mpsc_join
has been called on the channel object, the channel will close as soon as all producer thread callback functions have returned. In either case, the application defined consumer callback will receive a last message with its closed
argument set to true
, and the callback will no longer be called for that particular channel object.
It should be noted, however, that the call to mpsc_join
will hang, even when the channel closure has been requested by the consumer object using mpsc_consumer_close
, until all producer thread callback functions have returned. The mpsc_producer_ping
function exists so that a producer that is used as a worker can periodically check to make sure the channel is still opened, and, if not, return immediately. This is illustrated by both the the_first_wins.c and the proof_of_work.c examples.
Once the internal consumer thread and all the internal producer threads have been joined (which happens inside the mpsc_join
call), the mpsc_t
object will be cleaned up (i.e., destroyed), and the call to mpsc_join
will return, marking the end of the channel's life cycle.
For more details, you may have a look at the library's official API documentation website. I also encourage the interested reader to inspect the source code directly, by looking at src/mpsc.c.
- doxygen — A directory that contains Doxygen-related stuff used to generate the API documentation website for this library.
- examples — A directory that contains standalone examples illustrating how the library's different features can be used. The Makefile declares a recipe for each example. For instance, to run examples/quick_example.c simply run
make example_quick_example
(without the.c
extension at the end of the file name). - include — A directory that contains the header file mpsc.h; i.e., the declarations for the library's public API.
- src — A directory that contains the implementation file mpsc.c, in which all of the definitions for the functions and types declared in mpsc.h are provided.
- LICENSE — A file containing the license and copyright information for this project.
- Makefile — A
Makefile
(for use with GNU Make), which is provided as a convenience, and which can be used to automate operations such as building the library, running the examples, building the API documentation website, and installing/uninstalling the library on the target system. You may runmake
ormake help
for a list of all relevant recipes. WARNING: If you ever decide to usemake install
, please first make sure that/usr/local/{lib|include|man}
are valid installation paths on your system, and, if not, make sure to adjust them first. - VERSION — A simple text file that contains the library's current version. This is used by the Makefile to generate the documentation website and to "suffix" the library binaries with the current version number.
This is an experimental library, so please use with caution, and please feel free to let me know if you find anything wrong in my implementation.
- Currently, the library does not allow for asynchronous message sending; only synchronous message sending. As a consequence, each call to
mpsc_producer_send
(ormpsc_producer_send_empty
) will block until the previous internal call to the consumer callback (from inside the consumer thread callback loop) returns. I think that it would be interesting to add the option to create ampsc_t
instance with message buffering support, which would allow for asynchronous message passing. As far as the public API is concerned, that would require either modifying the return types (and possibly also the argument list) for bothmpsc_producer_send
andmpsc_producer_send_empty
, or simply creating new functions called, for instance,mpsc_producer_send_async
andmpsc_producer_send_empty_async
, to be used to send messages asynchronously. I think the second approach would be better since it would not break the existing API. - I would like to identify more real life applications for which this library could be useful and add those as examples to the examples directory.
- I want to add a section (to this README.md file) about error handling.
- I just realized, at the moment of publication of this first version of the project, that there is currently no way for a consumer callback to know for which
mpsc_t
object it is being executed (except than to use a callback for no more than only onempsc_t
instance by application). That means that, for an application using multiple channel instances, and for which knowledge about the channel object for which the callback is being executed would be required, that application would have to implement an individual consumer callback function for each channel, even if the callback implementation is the same across some of the channels. This could easily be solved by allowing to store arbitrary application data in thempsc_t
object at instantiation time, and by providing a simple function to be able to retrieve said data from inside the callback (e.g.,void *mpsc_consumer_channel_context(mpsc_consumer_t *self)
). A similar function to be used with the producer object could also be added (e.g.,void *mpsc_producer_channel_context(mpsc_producer_t *self)
).
If you have any questions, if you find bugs, or if you have suggestions for this project, please feel free to contact me by opening an issue on the repository.
This project is released under the MIT License.
Copyright (c) 2024 BB-301 ([email protected])