nilesr / cr Goto Github PK

Coroutines TS has been officially incorporated into C++21. Unfortunately it's not 2021 (yet), and also unfortunately, many projects use C not C++.

cr adds coroutines to C, with a round-robin scheduling system. You start a cr environment (cr_env) by providing a function to call and the number of frames, and cr creates that many frames and stacks, switching between them when the user-supplied function calls cr_yield(env). When you create an environment, you can optionally pass an array of n user data pointers, called batons, which will be passed to the corresponding frame. Each frame returns a single result, also a void pointer, which can be accessed through env->results after all the frames are done executing.

Here is an example:

#include <stdio.h>
#include "cr.h"

void* my_coroutine(cr_env* env, int frame_id, void* baton) {
   printf("Frame %d first time!\n", frame_id);
   cr_yield(env);
   printf("Frame %d second time around!\n", frame_id);
   return NULL;
}

int main(int argc, char** argv) {
   cr_env* env = cr_env_new(4);
   cr_run(env, &my_coroutine, NULL);
   printf("All frames have returned\n");
   cr_env_destroy(env);
}

This prints out the following

Frame 0 first time!
Frame 1 first time!
Frame 2 first time!
Frame 3 first time!
Frame 0 second time around!
Frame 1 second time around!
Frame 2 second time around!
Frame 3 second time around!
All frames have returned

cr also supports async IO. Here is an example cat program with no error handling that reads from stdin (0) and writes to stdout (1), and also allows other frames to do work while it is reading and writing

uint8_t buf[4096];
cr_aio_buf* buffer = cr_aio_buf_new(env, buf, 4096);
int nread;
while ((nread = cr_aio_read(buffer, 0)) > 0) {
	cr_aio_write(buffer, 1, nread);
}
cr_aio_buf_destroy(buffer);

The async IO functions properly return the number of bytes read or written, or -1 on error, and errno is set. They require linking against the POSIX realtime library, -lrt.

You can pass in NULL to cr_aio_buf_new if you would like a block of the passed size to be malloc'd instead.

You can build cr by running make in the current directory. This will also build any tests.

make outputs libcr.a, which can be linked against from your own project.

Creates a new cr environment with n frames. n must be positive. Returns a pointer to the environment.

The environment has the following helpful properties:

• int count: The n that was passed in to cr_env_new.
• int current: The frame ID of frame that is currently executing, or -1 if no frame is executing.
• int dead: The number of frames which have completed their execution and returned.
• void** results: An array of the results of each frame. Initialized to all NULLs before any of the frames have completed.
• cr_context* frames: The only thing interesting in here is the dead flag on each of the frames. If env->frames[i].dead is true for some i, then env->results[i] is safe to access. If env->dead == env->count then they are all safe to access.

cr does not expect any of these properties to be modified.

This starts all env->count frames executing func, where func has the following declaration:

void* (*cr_thread_function)(cr_env* env, int frame_id, void* baton);

Each frame will start executing the passed function at the beginning, with env being a pointer to the environment. The other two arguments are provided for convenience only, frame_id can be accessed from inside a frame by looking at env->current, and if batons were passed to cr_run then the baton is available in env->batons[frame_id].

When all frames have returned, their results are stored in env->results. The return value of this function is provided for convenience, it is also accessible through env->results.

Frees all resources associated with env. If you call this while frames are still executing, you're asking for serious trouble.

Call this function from inside a frame. It will save the state of the process into the current frame, and then find the next available frame and switch into it. If all other frames have returned, this may return immediately.

A crude version of await can be implemented using this, with this psuedocode

while (the thing is not ready yet) {
	cr_yield(env);
}
// the thing is ready

In particular, this is how the cr_aio functions work.

The following functions are global symbols because they must be called from files other than where they were defined, such as assembly files, but are meant for internal use only:

• cr_yield_do
• cr_run_internal
• cr_handle_result
• cr_handle_result_inner

Allocates metadata for a cr async IO buffer, including a pointer to the cr environment so it can cr_yield, a pointer to the buffer to hold data in, and the size of the buffer.

If the passed buf is NULL, a buffer of size is malloc'd, and will be free'd on cr_aio_buf_destroy

Performs an asynchronous read from the given file descriptor into the passed buffer.

While the read is in progress, it calls cr_yield to allow other frames to do work, but if the read happens fast enough, it may not yield.

If the read is successful, it returns the number of bytes read. If the read fails, this function returns -1 and errno is set appropriately.

Reads always begin from buf->start. buf->start starts at zero, and every time cr_aio_buf_write or cr_aio_buf_read returns a positive value, buf->start has been incremented by that value. To change the starting location of the read, just set buf->start appropriately before calling cr_aio_buf_read

Performs an asynchronous write of n bytes to the given file descriptor from the passed buffer. If n is greater than buf->size, this function returns -1 and errno is set to EDOM

While the write is in progress, it calls cr_yield to allow other frames to do work, but if the write happens fast enough, it may not yield.

If the write is successful, it returns the number of bytes written. If the write fails, this function returns -1 and errno is set appropriately.

Writes always begin from buf->start. buf->start starts at zero, and every time cr_aio_buf_write or cr_aio_buf_read returns a positive value, buf->start has been incremented by that value. To change the starting location of the write, just set buf->start appropriately before calling cr_aio_buf_write

Frees the resources associated with buf. If no backing buffer was pased to cr_aio_buf_new, the temporary buffer is also free'd.