避免使用 GCD Global队列创建Runloop常驻线程
本文对应 Demo 以及 Markdown 文件在仓库中,文中的错误可以提 PR 到这个文件,我会及时更改。
目录
GCD Global队列创建线程进行耗时操作的风险
先思考下如下几个问题:
- 新建线程的方式有哪些?各自的优缺点是什么?
- dispatch_async 函数分发到全局队列一定会新建线程执行任务么?
- 如果全局队列对应的线程池如果满了,后续的派发的任务会怎么处置?有什么风险?
答案大致是这样的:dispatch_async 函数分发到全局队列不一定会新建线程执行任务,全局队列底层有一个的线程池,如果线程池满了,那么后续的任务会被 block 住,等待前面的任务执行完成,才会继续执行。如果线程池中的线程长时间不结束,后续堆积的任务会越来越多,此时就会存在 APP crash的风险。
比如:
- (void)dispatchTest1 {
for (NSInteger i = 0; i< 10000 ; i++) {
dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
[self dispatchTask:i];
});
}
}
- (void)dispatchTask:(NSInteger)index {
//模拟耗时操作,比如DB,网络,文件读写等等
sleep(30);
NSLog(@"----:%ld",index);
}
以上逻辑用真机测试会有卡死的几率,并非每次都会发生,但多尝试几次就会复现,伴随前后台切换,crash几率增大。
下面做一下分析:
参看 GCD 源码我们可以看到全局队列的相关源码如下:
DISPATCH_NOINLINE
static void
_dispatch_queue_wakeup_global_slow(dispatch_queue_t dq, unsigned int n)
{
dispatch_root_queue_context_t qc = dq->do_ctxt;
uint32_t i = n;
int r;
_dispatch_debug_root_queue(dq, __func__);
dispatch_once_f(&_dispatch_root_queues_pred, NULL,
_dispatch_root_queues_init);
#if HAVE_PTHREAD_WORKQUEUES
#if DISPATCH_USE_PTHREAD_POOL
if (qc->dgq_kworkqueue != (void*)(~0ul))
#endif
{
_dispatch_root_queue_debug("requesting new worker thread for global "
"queue: %p", dq);
#if DISPATCH_USE_LEGACY_WORKQUEUE_FALLBACK
if (qc->dgq_kworkqueue) {
pthread_workitem_handle_t wh;
unsigned int gen_cnt;
do {
r = pthread_workqueue_additem_np(qc->dgq_kworkqueue,
_dispatch_worker_thread4, dq, &wh, &gen_cnt);
(void)dispatch_assume_zero(r);
} while (--i);
return;
}
#endif // DISPATCH_USE_LEGACY_WORKQUEUE_FALLBACK
#if HAVE_PTHREAD_WORKQUEUE_SETDISPATCH_NP
if (!dq->dq_priority) {
r = pthread_workqueue_addthreads_np(qc->dgq_wq_priority,
qc->dgq_wq_options, (int)i);
(void)dispatch_assume_zero(r);
return;
}
#endif
#if HAVE_PTHREAD_WORKQUEUE_QOS
r = _pthread_workqueue_addthreads((int)i, dq->dq_priority);
(void)dispatch_assume_zero(r);
#endif
return;
}
#endif // HAVE_PTHREAD_WORKQUEUES
#if DISPATCH_USE_PTHREAD_POOL
dispatch_pthread_root_queue_context_t pqc = qc->dgq_ctxt;
if (fastpath(pqc->dpq_thread_mediator.do_vtable)) {
while (dispatch_semaphore_signal(&pqc->dpq_thread_mediator)) {
if (!--i) {
return;
}
}
}
uint32_t j, t_count;
// seq_cst with atomic store to tail <rdar://problem/16932833>
t_count = dispatch_atomic_load2o(qc, dgq_thread_pool_size, seq_cst);
do {
if (!t_count) {
_dispatch_root_queue_debug("pthread pool is full for root queue: "
"%p", dq);
return;
}
j = i > t_count ? t_count : i;
} while (!dispatch_atomic_cmpxchgvw2o(qc, dgq_thread_pool_size, t_count,
t_count - j, &t_count, acquire));
pthread_attr_t *attr = &pqc->dpq_thread_attr;
pthread_t tid, *pthr = &tid;
#if DISPATCH_ENABLE_PTHREAD_ROOT_QUEUES
if (slowpath(dq == &_dispatch_mgr_root_queue)) {
pthr = _dispatch_mgr_root_queue_init();
}
#endif
do {
_dispatch_retain(dq);
while ((r = pthread_create(pthr, attr, _dispatch_worker_thread, dq))) {
if (r != EAGAIN) {
(void)dispatch_assume_zero(r);
}
_dispatch_temporary_resource_shortage();
}
} while (--j);
#endif // DISPATCH_USE_PTHREAD_POOL
}
对于执行的任务来说,所执行的线程具体是哪个线程,则是通过 GCD 的线程池(Thread Pool)来进行调度,正如Concurrent Programming: APIs and Challenges文章里给的示意图所示:
上面贴的源码,我们关注如下的部分:
其中有一个用来记录线程池大小的字段 dgq_thread_pool_size
。这个字段标记着GCD线程池的大小。摘录上面源码的一部分:
uint32_t j, t_count;
// seq_cst with atomic store to tail <rdar://problem/16932833>
t_count = dispatch_atomic_load2o(qc, dgq_thread_pool_size, seq_cst);
do {
if (!t_count) {
_dispatch_root_queue_debug("pthread pool is full for root queue: "
"%p", dq);
return;
}
j = i > t_count ? t_count : i;
} while (!dispatch_atomic_cmpxchgvw2o(qc, dgq_thread_pool_size, t_count,
t_count - j, &t_count, acquire));
从源码中我们可以对应到官方文档 :Getting the Global Concurrent Dispatch Queues里的说法:
A concurrent dispatch queue is useful when you have multiple tasks that can run in parallel. A concurrent queue is still a queue in that it dequeues tasks in a first-in, first-out order; however, a concurrent queue may dequeue additional tasks before any previous tasks finish. The actual number of tasks executed by a concurrent queue at any given moment is variable and can change dynamically as conditions in your application change. Many factors affect the number of tasks executed by the concurrent queues, including the number of available cores, the amount of work being done by other processes, and the number and priority of tasks in other serial dispatch queues.
也就是说:
全局队列的底层是一个线程池,向全局队列中提交的 block,都会被放到这个线程池中执行,如果线程池已满,后续再提交 block 就不会再重新创建线程。这就是为什么 Demo 会造成卡顿甚至冻屏的原因。
避免使用 GCD Global 队列创建 Runloop 常驻线程
在做网路请求时我们常常创建一个 Runloop 常驻线程用来接收、响应后续的服务端回执,比如NSURLConnection、AFNetworking等等,我们可以称这种线程为 Runloop 常驻线程。
正如上文所述,用 GCD Global 队列创建线程进行耗时操作是存在风险的。那么我们可以试想下,如果这个耗时操作变成了 runloop 常驻线程,会是什么结果?下面做一下分析:
先介绍下 Runloop 常驻线程的原理,在开发中一般有两种用法:
- 单一 Runloop 常驻线程:在 APP 的生命周期中开启了唯一的常驻线程来进行网络请求,常用于网络库,或者有维持长连接需求的库,比如: AFNetworking 、 SocketRocket。
- 多个 Runloop 常驻线程:每进行一次网络请求就开启一条 Runloop 常驻线程,这条线程的生命周期的起点是网络请求开始,终点是网络请求结束,或者网络请求超时。
单一 Runloop 常驻线程
先说第一种用法:
以 AFNetworking 为例,AFURLConnectionOperation 这个类是基于 NSURLConnection 构建的,其希望能在后台线程接收 Delegate 回调。为此 AFNetworking 单独创建了一个线程,并在这个线程中启动了一个 RunLoop:
+ (void)networkRequestThreadEntryPoint:(id)__unused object {
@autoreleasepool {
[[NSThread currentThread] setName:@"AFNetworking"];
NSRunLoop *runLoop = [NSRunLoop currentRunLoop];
[runLoop addPort:[NSMachPort port] forMode:NSDefaultRunLoopMode];
[runLoop run];
}
}
+ (NSThread *)networkRequestThread {
static NSThread *_networkRequestThread = nil;
static dispatch_once_t oncePredicate;
dispatch_once(&oncePredicate, ^{
_networkRequestThread = [[NSThread alloc] initWithTarget:self selector:@selector(networkRequestThreadEntryPoint:) object:nil];
[_networkRequestThread start];
});
return _networkRequestThread;
}
多个 Runloop 常驻线程
第二种用法,我写了一个小 Demo 来模拟这种场景,
我们模拟了一个场景:假设所有的网络请求全部超时,或者服务端根本不响应,然后网络库超时检测机制的做法:
#import "Foo.h"
@interface Foo() {
NSRunLoop *_runloop;
NSTimer *_timeoutTimer;
NSTimeInterval _timeoutInterval;
dispatch_semaphore_t _sem;
}
@end
@implementation Foo
- (instancetype)init {
if (!(self = [super init])) {
return nil;
}
_timeoutInterval = 1 ;
_sem = dispatch_semaphore_create(0);
// Do any additional setup after loading the view, typically from a nib.
return self;
}
- (id)test {
// 第一种方式:
// NSThread *networkRequestThread = [[NSThread alloc] initWithTarget:self selector:@selector(networkRequestThreadEntryPoint0:) object:nil];
// [networkRequestThread start];
//第二种方式:
dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^(void) {
[self networkRequestThreadEntryPoint0:nil];
});
dispatch_semaphore_wait(_sem, DISPATCH_TIME_FOREVER);
return @(YES);
}
- (void)networkRequestThreadEntryPoint0:(id)__unused object {
@autoreleasepool {
[[NSThread currentThread] setName:@"CYLTest"];
_runloop = [NSRunLoop currentRunLoop];
[_runloop addPort:[NSMachPort port] forMode:NSDefaultRunLoopMode];
_timeoutTimer = [NSTimer scheduledTimerWithTimeInterval:1 target:self selector:@selector(stopLoop) userInfo:nil repeats:NO];
[_runloop addTimer:_timeoutTimer forMode:NSRunLoopCommonModes];
[_runloop run];//在实际开发中最好使用这种方式来确保能runloop退出,做双重的保障[runloop runUntilDate:[NSDate dateWithTimeIntervalSinceNow:(timeoutInterval+5)]];
}
}
- (void)stopLoop {
CFRunLoopStop([_runloop getCFRunLoop]);
dispatch_semaphore_signal(_sem);
}
@end
如果
for (int i = 0; i < 300 ; i++) {
dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^(void) {
[[Foo new] test];
NSLog(@"