pthread_cond_wait deadlock in fifo circular queue - c

Question

my code is only using in one producer-one consumer situation.
here is my test code:
static void *afunc(void * arg) {
Queue* q = arg;
for(int i= 0; i< 100000; i++) {
*queue_pull(q) = i; //get one element space
queue_push(q); //increase the write pointer
}
return NULL;
}
static void *bfunc(void * arg) {
Queue* q = arg;
for(;;) {
int *i = queue_fetch(q); //get the first element in queue
printf("%d\n", *i);
queue_pop(q); //increase the read pointer
}
}
int main() {
Queue queue;
pthread_t a, b;
queue_init(&queue);
pthread_create(&a, NULL, afunc, &queue);
pthread_create(&b, NULL, bfunc, &queue);
sleep(100000);
return 0;
}
and here is the implementation of the circular queue
#define MAX_QUEUE_SIZE 3
typedef struct Queue{
int data[MAX_QUEUE_SIZE] ;
int read,write;
pthread_mutex_t mutex, mutex2;
pthread_cond_t not_empty, not_full;
}Queue;
int queue_init(Queue *queue) {
memset(queue, 0, sizeof(Queue));
pthread_mutex_init(&queue->mutex, NULL);
pthread_cond_init(&queue->not_empty, NULL);
pthread_mutex_init(&queue->mutex2, NULL);
pthread_cond_init(&queue->not_full, NULL);
return 0;
}
int* queue_fetch(Queue *queue) {
int* ret;
if (queue->read == queue->write) {
pthread_mutex_lock(&queue->mutex);
pthread_cond_wait(&queue->not_empty, &queue->mutex);
pthread_mutex_unlock(&queue->mutex);
}
ret = &(queue->data[queue->read]);
return ret;
}
void queue_pop(Queue *queue) {
nx_atomic_set(queue->read, (queue->read+1)%MAX_QUEUE_SIZE);
pthread_cond_signal(&queue->not_full);
}
int* queue_pull(Queue *queue) {
int* ret;
if ((queue->write+1)%MAX_QUEUE_SIZE == queue->read) {
pthread_mutex_lock(&queue->mutex2);
pthread_cond_wait(&queue->not_full, &queue->mutex2);
pthread_mutex_unlock(&queue->mutex2);
}
ret = &(queue->data[queue->write]);
return ret;
}
void queue_push(Queue *queue) {
nx_atomic_set(queue->write, (queue->write+1)%MAX_QUEUE_SIZE);
pthread_cond_signal(&queue->not_empty);
}
after a few moments, it seems the two child threads will turn into deadlock..
EDIT: i use two semaphore， but it also has some problem.. it's pretty
weird, if if just execute ./main, it seems fine, but if i redirect into a file, like ./main > a.txt, then wc -l a.txt, the result is not equal the enqueue number..
int queue_init(Queue *queue) {
memset(queue, 0, sizeof(Queue));
pthread_mutex_init(&queue->mutex, NULL);
sem_unlink("/not_empty");
queue->not_empty = sem_open("/not_empty", O_CREAT, 644, 0);
sem_unlink("/not_full");
queue->not_full = sem_open("/not_full", O_CREAT, 644, MAX_QUEUE_SIZE);
return 0;
}
int* queue_fetch(Queue *queue) {
sem_wait(queue->not_empty);
return &(queue->data[queue->read]);
}
void queue_pop(Queue *queue) {
nx_atomic_set(queue->read, (queue->read+1)%MAX_QUEUE_SIZE);
sem_post(queue->not_full);
}
int* queue_pull(Queue *queue) {
sem_wait(queue->not_full);
return &(queue->data[queue->write]);
}
void queue_push(Queue *queue) {
nx_atomic_set(queue->write, (queue->write+1)%MAX_QUEUE_SIZE);
sem_post(queue->not_empty);
}

Answer 1

You are manipulating the state of the queue outside the mutex, this is inherently racey.
I would suggest using a single mutex, but take it whenever you change or test the read & write indicies. This also means that you don't need the atomic sets.

Answer 2

Quite possibly one of your threads is waiting for a condition to be signalled after the signalling has occurred, causing both threads to wait for each other indefinitely.
Pthreads condition variables don't remain signalled -- the signalling is a momentary action. The condition variable isn't used determine whether to wait -- it's just used to wake up a thread that's already waiting; you need a different means for determining whether or not to wait, such as checking a flag or some sort of test condition.
Normally, you signal as follows:
Lock the mutex
Do your updates, generally leaving your test condition 'true' (eg. setting your flag)
Call pthread_cond_signal() or pthread_cond_broadcast()
Unlock the mutex
...and wait as follows:
Lock the mutex
Loop until your test expression is 'true' (eg. until your flag is set), calling pthread_cond_wait() only if the test is false (inside the loop).
After the loop, when your test has succeeded, do your work.
Unlock the mutex
For example, signalling might go something like this:
pthread_mutex_lock(&mtx); /* 1: lock mutex */
do_something_important(); /* 2: do your work... */
ready_flag = 1; /* ...and set the flag */
pthread_cond_signal(&cond); /* 3: signal the condition (before unlocking) */
pthread_mutex_unlock(&mtx); /* 4: unlock mutex */
and waiting might be something like this:
pthread_mutex_lock(&mtx); /* 1: lock mutex */
while (ready_flag == 0) /* 2: Loop until flag is set... */
pthread_cond_wait(&cond, &mtx); /* ...waiting when it isn't */
do_something_else(); /* 3: Do your work... */
ready_flag = 0; /* ...and clear the flag if it's all done */
pthread_mutex_unlock(&mtx); /* 4: unlock mutex */
The waiter won't miss the condition this way, because the mutex ensures that the waiter's test-and-wait and the signaller's set-and-signal cannot occur simultaneously.
This section of your queue_fetch() function:
if (queue->read == queue->write) {
pthread_mutex_lock(&queue->mutex);
pthread_cond_wait(&queue->not_empty, &queue->mutex);
pthread_mutex_unlock(&queue->mutex);
}
ret = &(queue->data[queue->read]);
..might be rewritten as follows:
pthread_mutex_lock(&queue->mutex);
while (queue->read == queue->write)
pthread_cond_wait(&queue->not_empty, &queue->mutex);
ret = &(queue->data[queue->read]);
pthread_mutex_unlock(&queue->mutex);
...where:
The lock/unlock of the mutex are moved around the if, so the mutex is held while the test expression is evaluated, and still held until the condition wait starts
The if is changed to a while in case the condition wait is prematurely interrupted
Access to queue->read and queue->write is done with the mutex held
Similar changes would be made to queue_pull().
As for the signalling code, the following section of queue_pop():
nx_atomic_set(queue->read, (queue->read+1)%MAX_QUEUE_SIZE);
pthread_cond_signal(&queue->not_full);
..might be changed to:
pthread_mutex_lock(&queue->mutex);
queue->read = (queue->read + 1) % MAX_QUEUE_SIZE;
pthread_cond_signal(&queue->not_full);
pthread_mutex_unlock(&queue->mutex);
..where:
The mutex is held while signalling the condition (this ensures the condition can't be signalled between the waiter deciding whether to wait and actually starting to wait, since the waiter would hold the mutex during that interval)
The mutex is held while changing queue->read as well rather than using nx_atomic_set() since the mutex is needed when signalling the condition anyway
Similar changes would be made to queue_push().
Additionally, you should just use a single mutex (so that the same mutex is always held when accessing read and write), and once the while loops are added to the condition waits there's little compelling reason to use more than one condition variable. If switching to a single condition variable, just signal the condition again after completing a wait:
pthread_mutex_lock(&queue->mutex);
while (queue->read == queue->write) {
pthread_cond_wait(&queue->cond, &queue->mutex);
pthread_cond_signal(&queue->cond); /* <-- signal next waiter, if any */
}
ret = &(queue->data[queue->read]);
pthread_mutex_unlock(&queue->mutex);