lpereira
commented
3 years ago Is the memory allocation for a coroutine struct+stack really worth optimizing? At least in Lwan, it didn't matter when I tried.
I took a look at your mempool implementation, and it does locking and whatnot -- might be a point of contention in multithreaded scenarios. You might want to try a different approach; for instance, using a thread-local ring buffer of pointers to memory (Lwan has an implementation you can use), where the algorithm would be the following pseudo-code:
pool = thread_local(mempool_new(sizeof(struct coro) + coro_stack_size)))
def pool_get(pool):
if pool.ring.is_empty():
return malloc(pool.size)
return pool.ring.get()
def pool_put(pool, mem):
if pool.ring.is_full():
free(mem)
else:
pool.ring.put(mem)No locking necessary, etc.
In any case, I don't know what problems you're referring to. Is the program crashing? If so, I'd try running under Valgrind and/or Address Sanitizer first.
jojo0severo
So, i'm currently trying to use your coroutines as a standalone, extracting the code was not the problem but the attempts to optmize it were.
I built a memory pool to pre allocate memory for new coroutines Replacing malloc and free on:
coro_new:
struct coro *coro = malloc(sizeof(*coro) + CORO_STACK_MIN);bystruct coro *coro = alloc_mem(pool);coro_free:
free(coro)byfree_mem(pool, coro)Pool is correctly implemented as the problems only happen when using it with the coroutines.
Since I'm not using the defer function, is there any other location where i should replace malloc and free?
Thanks for any reply.
Coroutines with memory pool: https://github.com/jojo0severo/lwan-coro-with-mempool