深入理解GCD之dispatch_group

NeroXie發表於2019-02-27
GC

原文連結深入理解GCD之dispatch_group

之前已經介紹了dispatch_semaphore的底層實現，dispatch_group的實現是基於前者的。在看原始碼之前，我們先看一下我們是如何應用的。假設有這麼場景：有一個A耗時操作，B和C兩個網路請求和一個耗時操作C當ABC都執行完成後，重新整理頁面。我們可以用dispatch_group實現。關鍵如下：

- (void)viewDidLoad {
    [super viewDidLoad];
    
        __block NSInteger number = 0;
    
    dispatch_group_t group = dispatch_group_create();
    
    //A耗時操作
    dispatch_group_async(group, dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
        sleep(3);
        number += 2222;
    });
    
    //B網路請求
    dispatch_group_enter(group);
    [self sendRequestWithCompletion:^(id response) {
        number += [response integerValue];
        dispatch_group_leave(group);
    }];
    
    //C網路請求
    dispatch_group_enter(group);
    [self sendRequestWithCompletion:^(id response) {
        number += [response integerValue];
        dispatch_group_leave(group);
    }];
    
    dispatch_group_notify(group, dispatch_get_main_queue(), ^{
        NSLog(@"%zd", number);
    });
}

- (void)sendRequestWithCompletion:(void (^)(id response))completion {
    //模擬一個網路請求
    dispatch_queue_t queue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0);
    dispatch_async(queue, ^{
        sleep(2);
        dispatch_async(dispatch_get_main_queue(), ^{
            if (completion) completion(@1111);
        });
    });
}
複製程式碼

接下來我們根據上面的流程來看一下dispatch_group的相關API

dispatch_group_create

dispatch_group_t
dispatch_group_create(void)
{
	return (dispatch_group_t)dispatch_semaphore_create(LONG_MAX);
}
複製程式碼

dispatch_group_create其實就是建立了一個value為LONG_MAX的dispatch_semaphore訊號量

dispatch_group_async

void
dispatch_group_async(dispatch_group_t dg, dispatch_queue_t dq,
		dispatch_block_t db)
{
	dispatch_group_async_f(dg, dq, _dispatch_Block_copy(db),
			_dispatch_call_block_and_release);
}
複製程式碼

dispatch_group_async只是dispatch_group_async_f的封裝

dispatch_group_async_f

void
dispatch_group_async_f(dispatch_group_t dg, dispatch_queue_t dq, void *ctxt,
		dispatch_function_t func)
{
	dispatch_continuation_t dc;

	_dispatch_retain(dg);
	dispatch_group_enter(dg);

	dc = fastpath(_dispatch_continuation_alloc_cacheonly());
	if (!dc) {
		dc = _dispatch_continuation_alloc_from_heap();
	}

	dc->do_vtable = (void *)(DISPATCH_OBJ_ASYNC_BIT | DISPATCH_OBJ_GROUP_BIT);
	dc->dc_func = func;
	dc->dc_ctxt = ctxt;
	dc->dc_group = dg;

	// No fastpath/slowpath hint because we simply don't know
	if (dq->dq_width != 1 && dq->do_targetq) {
		return _dispatch_async_f2(dq, dc);
	}

	_dispatch_queue_push(dq, dc);
}
複製程式碼

從上面的程式碼我們可以看出dispatch_group_async_f和dispatch_async_f相似。dispatch_group_async_f多了dispatch_group_enter(dg);，另外在do_vtable的賦值中dispatch_group_async_f多了一個DISPATCH_OBJ_GROUP_BIT的標記符。既然新增了dispatch_group_enter必定會存在dispatch_group_leave。在之前《深入理解GCD之dispatch_queue》介紹_dispatch_continuation_pop函式的原始碼中有一段程式碼如下：

	_dispatch_client_callout(dc->dc_ctxt, dc->dc_func);
	if (dg) {
		//group需要進行呼叫dispatch_group_leave並釋放訊號
		dispatch_group_leave(dg);
		_dispatch_release(dg);
	}
複製程式碼

所以dispatch_group_async_f函式中的dispatch_group_leave是在_dispatch_continuation_pop函式中呼叫的。

這裡概括一下dispatch_group_async_f的工作流程：

  1. 呼叫dispatch_group_enter；
  2. 將block和queue等資訊記錄到dispatch_continuation_t結構體中，並將它加入到group的連結串列中；
  3. _dispatch_continuation_pop執行時會判斷任務是否為group，是的話執行完任務再呼叫dispatch_group_leave以達到訊號量的平衡。

dispatch_group_enter

void
dispatch_group_enter(dispatch_group_t dg)
{
	dispatch_semaphore_t dsema = (dispatch_semaphore_t)dg;

	(void)dispatch_semaphore_wait(dsema, DISPATCH_TIME_FOREVER);
}
複製程式碼

dispatch_group_enter將dispatch_group_t轉換成dispatch_semaphore_t，並呼叫dispatch_semaphore_wait，原子性減1後，進入等待狀態直到有訊號喚醒。所以說dispatch_group_enter就是對dispatch_semaphore_wait的封裝。

dispatch_group_leave

void
dispatch_group_leave(dispatch_group_t dg)
{
	dispatch_semaphore_t dsema = (dispatch_semaphore_t)dg;
	dispatch_atomic_release_barrier();
	long value = dispatch_atomic_inc2o(dsema, dsema_value);//dsema_value原子性加1
	if (slowpath(value == LONG_MIN)) {//記憶體溢位，由於dispatch_group_leave在dispatch_group_enter之前呼叫
		DISPATCH_CLIENT_CRASH("Unbalanced call to dispatch_group_leave()");
	}
	if (slowpath(value == dsema->dsema_orig)) {//表示所有任務已經完成，喚醒group
		(void)_dispatch_group_wake(dsema);
	}
}
複製程式碼

從上面的原始碼中我們看到dispatch_group_leave將dispatch_group_t轉換成dispatch_semaphore_t後將dsema_value的值原子性加1。如果value為LONG_MIN程式crash；如果value等於dsema_orig表示所有任務已完成，呼叫_dispatch_group_wake喚醒group（_dispatch_group_wake的用於和notify有關，我們會在後面介紹）。因為在enter的時候進行了原子性減1操作。所以在leave的時候需要原子性加1。

這裡先說明一下enter和leave之間的關係：

  1. dispatch_group_leave與dispatch_group_enter配對使用。當呼叫了dispatch_group_enter而沒有呼叫dispatch_group_leave時，由於value不等於dsema_orig不會走到喚醒邏輯，dispatch_group_notify中的任務無法執行或者dispatch_group_wait收不到訊號而卡住執行緒。

  2. dispatch_group_enter必須在dispatch_group_leave之前出現。當dispatch_group_leave比dispatch_group_enter多呼叫了一次或者說在dispatch_group_enter之前被呼叫的時候，dispatch_group_leave進行原子性加1操作，相當於value為LONGMAX+1，發生資料長度溢位，變成LONG_MIN，由於value == LONG_MIN成立，程式發生crash。

dispatch_group_notify

void
dispatch_group_notify(dispatch_group_t dg, dispatch_queue_t dq,
		dispatch_block_t db)
{
	dispatch_group_notify_f(dg, dq, _dispatch_Block_copy(db),
			_dispatch_call_block_and_release);
}
複製程式碼

dispatch_group_notify是dispatch_group_notify_f的封裝，具體實現在後者。

dispatch_group_notify_f

void
dispatch_group_notify_f(dispatch_group_t dg, dispatch_queue_t dq, void *ctxt,
		void (*func)(void *))
{
	dispatch_semaphore_t dsema = (dispatch_semaphore_t)dg;
	struct dispatch_sema_notify_s *dsn, *prev;

	//封裝dispatch_continuation_t結構體
	// FIXME -- this should be updated to use the continuation cache
	while (!(dsn = calloc(1, sizeof(*dsn)))) {
		sleep(1);
	}

	dsn->dsn_queue = dq;
	dsn->dsn_ctxt = ctxt;
	dsn->dsn_func = func;
	_dispatch_retain(dq);
	dispatch_atomic_store_barrier();
	//將結構體放到連結串列尾部，如果連結串列為空同時設定連結串列頭部節點並喚醒group
	prev = dispatch_atomic_xchg2o(dsema, dsema_notify_tail, dsn);
	if (fastpath(prev)) {
		prev->dsn_next = dsn;
	} else {
		_dispatch_retain(dg);
		(void)dispatch_atomic_xchg2o(dsema, dsema_notify_head, dsn);
		if (dsema->dsema_value == dsema->dsema_orig) {//任務已經完成，喚醒group
			_dispatch_group_wake(dsema);
		}
	}
}
複製程式碼

所以dispatch_group_notify函式只是用連結串列把所有回撥通知儲存起來，等待呼叫。

_dispatch_group_wake

static long
_dispatch_group_wake(dispatch_semaphore_t dsema)
{
	struct dispatch_sema_notify_s *next, *head, *tail = NULL;
	long rval;
	//將dsema的dsema_notify_head賦值為NULL，同時將之前的內容賦給head
	head = dispatch_atomic_xchg2o(dsema, dsema_notify_head, NULL);
	if (head) {
		// snapshot before anything is notified/woken <rdar://problem/8554546>
		//將dsema的dsema_notify_tail賦值為NULL，同時將之前的內容賦給tail
		tail = dispatch_atomic_xchg2o(dsema, dsema_notify_tail, NULL);
	}
	//將dsema的dsema_group_waiters設定為0，並返回原來的值
	rval = dispatch_atomic_xchg2o(dsema, dsema_group_waiters, 0);
	if (rval) {
		//迴圈呼叫semaphore_signal喚醒當初等待group的訊號量，使得dispatch_group_wait函式返回。
		// wake group waiters
#if USE_MACH_SEM
		_dispatch_semaphore_create_port(&dsema->dsema_waiter_port);
		do {
			kern_return_t kr = semaphore_signal(dsema->dsema_waiter_port);
			DISPATCH_SEMAPHORE_VERIFY_KR(kr);
		} while (--rval);
#elif USE_POSIX_SEM
		do {
			int ret = sem_post(&dsema->dsema_sem);
			DISPATCH_SEMAPHORE_VERIFY_RET(ret);
		} while (--rval);
#endif
	}
	if (head) {
		//獲取連結串列，依次呼叫dispatch_async_f非同步執行在notify函式中的任務即Block。
		// async group notify blocks
		do {
			dispatch_async_f(head->dsn_queue, head->dsn_ctxt, head->dsn_func);
			_dispatch_release(head->dsn_queue);
			next = fastpath(head->dsn_next);
			if (!next && head != tail) {
				while (!(next = fastpath(head->dsn_next))) {
					_dispatch_hardware_pause();
				}
			}
			free(head);
		} while ((head = next));
		_dispatch_release(dsema);
	}
	return 0;
}
複製程式碼

_dispatch_group_wake主要的作用有兩個：

  1. 呼叫semaphore_signal喚醒當初等待group的訊號量，使得dispatch_group_wait函式返回。

  2. 獲取連結串列，依次呼叫dispatch_async_f非同步執行在notify函式中的任務即Block。

到這裡我們已經差不多知道了dispatch_group工作過程，我們用一張圖表示：

dispatch_group

dispatch_group_wait

long
dispatch_group_wait(dispatch_group_t dg, dispatch_time_t timeout)
{
	dispatch_semaphore_t dsema = (dispatch_semaphore_t)dg;

	if (dsema->dsema_value == dsema->dsema_orig) {//沒有需要執行的任務
		return 0;
	}
	if (timeout == 0) {//返回超時
#if USE_MACH_SEM
		return KERN_OPERATION_TIMED_OUT;
#elif USE_POSIX_SEM
		errno = ETIMEDOUT;
		return (-1);
#endif
	}
	return _dispatch_group_wait_slow(dsema, timeout);
}
複製程式碼

dispatch_group_wait用於等待group中的任務完成。

_dispatch_group_wait_slow

static long
_dispatch_group_wait_slow(dispatch_semaphore_t dsema, dispatch_time_t timeout)
{
	long orig;

again:
	// check before we cause another signal to be sent by incrementing
	// dsema->dsema_group_waiters
	if (dsema->dsema_value == dsema->dsema_orig) {
		return _dispatch_group_wake(dsema);
	}
	// Mach semaphores appear to sometimes spuriously wake up. Therefore,
	// we keep a parallel count of the number of times a Mach semaphore is
	// signaled (6880961).
	(void)dispatch_atomic_inc2o(dsema, dsema_group_waiters);
	// check the values again in case we need to wake any threads
	if (dsema->dsema_value == dsema->dsema_orig) {
		return _dispatch_group_wake(dsema);
	}

#if USE_MACH_SEM
	mach_timespec_t _timeout;
	kern_return_t kr;

	_dispatch_semaphore_create_port(&dsema->dsema_waiter_port);

	// From xnu/osfmk/kern/sync_sema.c:
	// wait_semaphore->count = -1; /* we don't keep an actual count */
	//
	// The code above does not match the documentation, and that fact is
	// not surprising. The documented semantics are clumsy to use in any
	// practical way. The above hack effectively tricks the rest of the
	// Mach semaphore logic to behave like the libdispatch algorithm.

	switch (timeout) {
	default:
		do {
			uint64_t nsec = _dispatch_timeout(timeout);
			_timeout.tv_sec = (typeof(_timeout.tv_sec))(nsec / NSEC_PER_SEC);
			_timeout.tv_nsec = (typeof(_timeout.tv_nsec))(nsec % NSEC_PER_SEC);
			kr = slowpath(semaphore_timedwait(dsema->dsema_waiter_port,
					_timeout));
		} while (kr == KERN_ABORTED);

		if (kr != KERN_OPERATION_TIMED_OUT) {
			DISPATCH_SEMAPHORE_VERIFY_KR(kr);
			break;
		}
		// Fall through and try to undo the earlier change to
		// dsema->dsema_group_waiters
	case DISPATCH_TIME_NOW:
		while ((orig = dsema->dsema_group_waiters)) {
			if (dispatch_atomic_cmpxchg2o(dsema, dsema_group_waiters, orig,
					orig - 1)) {
				return KERN_OPERATION_TIMED_OUT;
			}
		}
		// Another thread called semaphore_signal().
		// Fall through and drain the wakeup.
	case DISPATCH_TIME_FOREVER:
		do {
			kr = semaphore_wait(dsema->dsema_waiter_port);
		} while (kr == KERN_ABORTED);
		DISPATCH_SEMAPHORE_VERIFY_KR(kr);
		break;
	}
#elif USE_POSIX_SEM
//這部分程式碼省略
#endif

	goto again;
}
複製程式碼

從上面的程式碼我們發現_dispatch_group_wait_slow和_dispatch_semaphore_wait_slow的邏輯很接近。都利用mach核心的semaphore進行訊號的傳送。區別在於_dispatch_semaphore_wait_slow在等待結束後是return，而_dispatch_group_wait_slow在等待結束是呼叫_dispatch_group_wake去喚醒這個group。

總結

  1. dispatch_group是一個初始值為LONG_MAX的訊號量，group中的任務完成是判斷其value是否恢復成初始值。

  2. dispatch_group_enter和dispatch_group_leave必須成對使用並且支援巢狀。

  3. 如果dispatch_group_enter比dispatch_group_leave多，由於value不等於dsema_orig不會走到喚醒邏輯，dispatch_group_notify中的任務無法執行或者dispatch_group_wait收不到訊號而卡住執行緒。如果是dispatch_group_leave多，則會引起崩潰。

相關文章