| 1 | #include "pthread_impl.h" |
| 2 | |
| 3 | static int pshared_barrier_wait(pthread_barrier_t *b) |
| 4 | { |
| 5 | 	int limit = (b->_b_limit & INT_MAX) + 1; |
| 6 | 	int ret = 0; |
| 7 | 	int v, w; |
| 8 | |
| 9 | 	if (limit==1) return PTHREAD_BARRIER_SERIAL_THREAD; |
| 10 | |
| 11 | 	while ((v=a_cas(&b->_b_lock, 0, limit))) |
| 12 | 		__wait(&b->_b_lock, &b->_b_waiters, v, 0); |
| 13 | |
| 14 | 	/* Wait for <limit> threads to get to the barrier */ |
| 15 | 	if (++b->_b_count == limit) { |
| 16 | 		a_store(&b->_b_count, 0); |
| 17 | 		ret = PTHREAD_BARRIER_SERIAL_THREAD; |
| 18 | 		if (b->_b_waiters2) __wake(&b->_b_count, -1, 0); |
| 19 | 	} else { |
| 20 | 		a_store(&b->_b_lock, 0); |
| 21 | 		if (b->_b_waiters) __wake(&b->_b_lock, 1, 0); |
| 22 | 		while ((v=b->_b_count)>0) |
| 23 | 			__wait(&b->_b_count, &b->_b_waiters2, v, 0); |
| 24 | 	} |
| 25 | |
| 26 | 	__vm_lock(); |
| 27 | |
| 28 | 	/* Ensure all threads have a vm lock before proceeding */ |
| 29 | 	if (a_fetch_add(&b->_b_count, -1)==1-limit) { |
| 30 | 		a_store(&b->_b_count, 0); |
| 31 | 		if (b->_b_waiters2) __wake(&b->_b_count, -1, 0); |
| 32 | 	} else { |
| 33 | 		while ((v=b->_b_count)) |
| 34 | 			__wait(&b->_b_count, &b->_b_waiters2, v, 0); |
| 35 | 	} |
| 36 | 	 |
| 37 | 	/* Perform a recursive unlock suitable for self-sync'd destruction */ |
| 38 | 	do { |
| 39 | 		v = b->_b_lock; |
| 40 | 		w = b->_b_waiters; |
| 41 | 	} while (a_cas(&b->_b_lock, v, v==INT_MIN+1 ? 0 : v-1) != v); |
| 42 | |
| 43 | 	/* Wake a thread waiting to reuse or destroy the barrier */ |
| 44 | 	if (v==INT_MIN+1 || (v==1 && w)) |
| 45 | 		__wake(&b->_b_lock, 1, 0); |
| 46 | |
| 47 | 	__vm_unlock(); |
| 48 | |
| 49 | 	return ret; |
| 50 | } |
| 51 | |
| 52 | struct instance |
| 53 | { |
| 54 | 	volatile int count; |
| 55 | 	volatile int last; |
| 56 | 	volatile int waiters; |
| 57 | 	volatile int finished; |
| 58 | }; |
| 59 | |
| 60 | int pthread_barrier_wait(pthread_barrier_t *b) |
| 61 | { |
| 62 | 	int limit = b->_b_limit; |
| 63 | 	struct instance *inst; |
| 64 | |
| 65 | 	/* Trivial case: count was set at 1 */ |
| 66 | 	if (!limit) return PTHREAD_BARRIER_SERIAL_THREAD; |
| 67 | |
| 68 | 	/* Process-shared barriers require a separate, inefficient wait */ |
| 69 | 	if (limit < 0) return pshared_barrier_wait(b); |
| 70 | |
| 71 | 	/* Otherwise we need a lock on the barrier object */ |
| 72 | 	while (a_swap(&b->_b_lock, 1)) |
| 73 | 		__wait(&b->_b_lock, &b->_b_waiters, 1, 1); |
| 74 | 	inst = b->_b_inst; |
| 75 | |
| 76 | 	/* First thread to enter the barrier becomes the "instance owner" */ |
| 77 | 	if (!inst) { |
| 78 | 		struct instance new_inst = { 0 }; |
| 79 | 		int spins = 200; |
| 80 | 		b->_b_inst = inst = &new_inst; |
| 81 | 		a_store(&b->_b_lock, 0); |
| 82 | 		if (b->_b_waiters) __wake(&b->_b_lock, 1, 1); |
| 83 | 		while (spins-- && !inst->finished) |
| 84 | 			a_spin(); |
| 85 | 		a_inc(&inst->finished); |
| 86 | 		while (inst->finished == 1) |
| 87 | 			__syscall(SYS_futex,&inst->finished,FUTEX_WAIT|FUTEX_PRIVATE,1,0) != -ENOSYS |
| 88 | 			|| __syscall(SYS_futex,&inst->finished,FUTEX_WAIT,1,0); |
| 89 | 		return PTHREAD_BARRIER_SERIAL_THREAD; |
| 90 | 	} |
| 91 | |
| 92 | 	/* Last thread to enter the barrier wakes all non-instance-owners */ |
| 93 | 	if (++inst->count == limit) { |
| 94 | 		b->_b_inst = 0; |
| 95 | 		a_store(&b->_b_lock, 0); |
| 96 | 		if (b->_b_waiters) __wake(&b->_b_lock, 1, 1); |
| 97 | 		a_store(&inst->last, 1); |
| 98 | 		if (inst->waiters) |
| 99 | 			__wake(&inst->last, -1, 1); |
| 100 | 	} else { |
| 101 | 		a_store(&b->_b_lock, 0); |
| 102 | 		if (b->_b_waiters) __wake(&b->_b_lock, 1, 1); |
| 103 | 		__wait(&inst->last, &inst->waiters, 0, 1); |
| 104 | 	} |
| 105 | |
| 106 | 	/* Last thread to exit the barrier wakes the instance owner */ |
| 107 | 	if (a_fetch_add(&inst->count,-1)==1 && a_fetch_add(&inst->finished,1)) |
| 108 | 		__wake(&inst->finished, 1, 1); |
| 109 | |
| 110 | 	return 0; |
| 111 | } |