被遗弃的线程

2018-04-14 本文已影响579人 sindri的小巢

image

main函数作为程序运行的入口，正常情况下，函数会执行毫秒级别的操作，然后返回一个0表示程序正常终止。为了避免应用启动即终止，苹果设计了runloop机制来维持线程的生命，runloop在每一次的循环当中不断的去处理事件，或控制线程的休眠和唤醒。runloop还结合了libdispatch的任务派发机制，可以循环地处理async到队列中的任务

启动runloop

从runloop对外暴露的接口来看，启动方式一共存在三种：

无条件的启动。这种启动方式缺少停止runloop的手段，唯一的结束方式是kill掉线程
```
  - (void)run;
```
设定超时时间。直接run会调用这个接口并且传入distantFuture，表示无限长的超时时间。如果超时时间不是无限长，那么runloop会在处理完事件或者超时后终止，优于直接run
```
  - (void)runUntilDate: (NSDate *)limitDate;
```
设置超时时间和运行模式。比起第二种，允许我们让runloop运行在某个模式下，灵活性更高
```
  - (BOOL)runMode: (NSRunLoopMode)mode beforeDate: (NSDate *)limitDate;
```

如果runloop在启动之后没有任何sources、timers或者ports事件可以处理，那么会自动退出，否则会在处理完成后让线程陷入休眠，等待这些事件重新唤醒线程处理。下面是最常用来表示runloop处理逻辑的示意图：

image

除开图中列出的事件之外，main loop会处理timer之后检测队列中是否存在待执行的block然后开始执行

什么情况下需要启动runloop

主线程的runloop会在应用启动后被UIApplication启动，其他线程则需要我们主动去run。从接触iOS开发到现在，笔者了解的需要主动启动runloop只有这么两类：

子线程使用timer

由于NSTimer本身就不是一个能确保稳定回调的定时器机制，并且主线程会经常处在忙碌状态，这又进一步降低了NSTimer的准确性。为了提高定时的准确性，多数人会采用子线程启动runloop的方式来实现定时器功能
线程保活

线程保活是一种不太常见的需求，但是如果你曾经了解过AFNetworking的做法，会发现创建了子线程之后，采用一个空port的方式来启动runloop，避免线程被中止回收。但实际上这种做法很容易导致线程既无法被回收，也不能被使用的情况

上面两种不同的应用场景，实际上是使用timer和port维持runloop不会因为没有事件处理直接退出，而且在这些源事件来临之前，线程大多数情况下处在休眠状态不造成额外损耗

串行队列的runloop

假设现在需要使用一个子线程的runloop来实现定时器，由于runloop在停止之前，线程会一直存活，因此可能会想利用这个存活的线程处理其他的任务。因此除了NSTimer之外，我们添加一个GCD Timer定时的派发任务给这个启动runloop的队列：

dispatch_queue_t serialQueue = dispatch_queue_create("serial.queue", DISPATCH_QUEUE_SERIAL);

dispatch_async(serialQueue, ^{
    NSLog(@"the task run in the thread: %d", mach_thread_self());
    [NSTimer scheduledTimerWithTimeInterval: 0.5 repeats: YES block: ^(NSTimer * _Nonnull timer) {
        NSLog(@"ns timer in the thread: %d", mach_thread_self());
    }];
    [[NSRunLoop currentRunLoop] runUntilDate: [NSDate dateWithTimeIntervalSinceNow: 600]];
});

dispatch_source_t timer = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, serialQueue);
dispatch_source_set_timer(timer, DISPATCH_TIME_NOW, 0.5 * NSEC_PER_SEC, 0 * NSEC_PER_SEC);
dispatch_source_set_event_handler(timer, ^{
    NSLog(@"gcd timer in the thread: %d", mach_thread_self());
});
dispatch_resume(timer);

按照预期，这段代码充分利用了已经被保活的线程，除了已有的NSTimer之外，线程还能在空闲的时间去处理不断派发的任务，但实际上只有NSTimer的任务被执行：

2018-04-14 10:32:54.667718+0800 PThreads[7693:94702] the task run in the thread: 4355
2018-04-14 10:32:55.168871+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:55.672011+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:56.169150+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:56.669411+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:57.169665+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:57.669234+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:58.172068+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:58.669446+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:59.169223+0800 PThreads[7693:94702] ns timer in the thread: 4355
2018-04-14 10:32:59.671802+0800 PThreads[7693:94702] ns timer in the thread: 4355

导致保活线程无法处理async任务的原因有两个：

runloop和queue的区别

runloop和queue各自维护着自己的一个任务队列，在runloop的每个周期里面，会检测自身的任务队列里面是否存在待执行的task并且执行。但主线程的情况比较特殊，在main runloop的每个周期，会去检测main queue是否存在待执行任务，如果存在，那么copy到自身的任务队列中执行
async的实现不同

在非主线程之外，runloop和queue的任务队列是互不干扰的，因此两者处理任务的机制也是完全不同的。当async任务到队列时，GCD会尝试寻找一个线程来执行任务。由于串行队列同时只能与一个线程挂钩，因此GCD会让该线程执行完已有任务后，才执行async到队列中的任务。但由于线程被保活，任务是一个条件死循环condition-loop，因此async的任务始终无法被处理

为了证明这些原因，可以通过CFRunLoopPerformBlock将任务直接加入到runloop自身的任务队列中，检测这个任务是否被执行：

__block CFRunLoopRef serialRunLoop = NULL;
dispatch_queue_t mainQueue = dispatch_get_main_queue();
dispatch_queue_t serialQueue = dispatch_queue_create("serial.queue", DISPATCH_QUEUE_SERIAL);

dispatch_async(serialQueue, ^{
    NSLog(@"the task run in the thread: %d", mach_thread_self());
    [NSTimer scheduledTimerWithTimeInterval: 0.5 repeats: YES block: ^(NSTimer * _Nonnull timer) {
        NSLog(@"ns timer in the thread: %d", mach_thread_self());
    }];
    serialRunLoop = [NSRunLoop currentRunLoop].getCFRunLoop;
    [[NSRunLoop currentRunLoop] runUntilDate: [NSDate dateWithTimeIntervalSinceNow: 600]];
});

dispatch_source_t timer = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, mainQueue);
dispatch_source_set_timer(timer, DISPATCH_TIME_NOW, 0.5 * NSEC_PER_SEC, 0 * NSEC_PER_SEC);
dispatch_source_set_event_handler(timer, ^{
    dispatch_async(serialQueue, ^{
        NSLog(@"gcd timer in the thread: %d", mach_thread_self());
    });
    CFRunLoopPerformBlock(serialRunLoop, NSDefaultRunLoopMode, ^{
        NSLog(@"perform block in thread: %d", mach_thread_self());
    });
});
dispatch_resume(timer);

再次运行之后，async的任务依旧无法被处理，但是perform block的任务总是能在timer唤醒休眠的线程后被处理：

2018-04-14 11:01:29.925924+0800 PThreads[15619:198121] the task run in the thread: 4355
2018-04-14 11:01:30.428175+0800 PThreads[15619:198121] ns timer in the thread: 4355
2018-04-14 11:01:30.428982+0800 PThreads[15619:198121] perform block in thread: 4355
2018-04-14 11:01:30.429410+0800 PThreads[15619:198121] perform block in thread: 4355
2018-04-14 11:01:30.932411+0800 PThreads[15619:198121] ns timer in the thread: 4355
2018-04-14 11:01:30.932674+0800 PThreads[15619:198121] perform block in thread: 4355
2018-04-14 11:01:31.430281+0800 PThreads[15619:198121] ns timer in the thread: 4355
2018-04-14 11:01:31.430546+0800 PThreads[15619:198121] perform block in thread: 4355
2018-04-14 11:01:31.929485+0800 PThreads[15619:198121] ns timer in the thread: 4355
2018-04-14 11:01:31.929691+0800 PThreads[15619:198121] perform block in thread: 4355
2018-04-14 11:01:32.432369+0800 PThreads[15619:198121] ns timer in the thread: 4355
2018-04-14 11:01:32.432726+0800 PThreads[15619:198121] perform block in thread: 4355
2018-04-14 11:01:32.930981+0800 PThreads[15619:198121] ns timer in the thread: 4355
2018-04-14 11:01:32.931179+0800 PThreads[15619:198121] perform block in thread: 4355
2018-04-14 11:01:33.429207+0800 PThreads[15619:198121] ns timer in the thread: 4355
2018-04-14 11:01:33.429519+0800 PThreads[15619:198121] perform block in thread: 4355

并行队列的runloop

队列的串并行属性决定了队列能不能被多个线程处理任务，因此同样的代码在并行队列执行，产生的结果必然是有所区别的：

__block CFRunLoopRef serialRunLoop = NULL;
dispatch_queue_t mainQueue = dispatch_get_main_queue();
dispatch_queue_t serialQueue = dispatch_queue_create("concurrent.queue", DISPATCH_QUEUE_CONCURRENT);

dispatch_async(serialQueue, ^{
    NSLog(@"the task run in the thread: %d", mach_thread_self());
    [NSTimer scheduledTimerWithTimeInterval: 0.5 repeats: YES block: ^(NSTimer * _Nonnull timer) {
        NSLog(@"ns timer in the thread: %d", mach_thread_self());
    }];
    serialRunLoop = [NSRunLoop currentRunLoop].getCFRunLoop;
    [[NSRunLoop currentRunLoop] runUntilDate: [NSDate dateWithTimeIntervalSinceNow: 600]];
});

dispatch_source_t timer = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, mainQueue);
dispatch_source_set_timer(timer, DISPATCH_TIME_NOW, 0.5 * NSEC_PER_SEC, 0 * NSEC_PER_SEC);
dispatch_source_set_event_handler(timer, ^{
    dispatch_async(serialQueue, ^{
        NSLog(@"gcd timer in the thread: %d", mach_thread_self());
    });
    CFRunLoopPerformBlock(serialRunLoop, NSDefaultRunLoopMode, ^{
        NSLog(@"perform block in thread: %d", mach_thread_self());
    });
});
dispatch_resume(timer);

输出结果如下：

2018-04-14 11:02:47.348083+0800 PThreads[15988:203200] the task run in the thread: 4867
2018-04-14 11:02:47.397174+0800 PThreads[15988:203203] gcd timer in the thread: 3843
2018-04-14 11:02:47.840678+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:47.852870+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:47.853122+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:47.853552+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:48.340602+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:48.352863+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:48.353149+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:48.840085+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:48.853918+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:48.854120+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:49.340729+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:49.354172+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:49.354470+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:49.840661+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:49.853115+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:49.853288+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:50.340078+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:50.354262+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:50.354537+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:50.840653+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:50.854117+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:50.854406+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:51.339972+0800 PThreads[15988:203206] gcd timer in the thread: 9219
2018-04-14 11:02:51.353246+0800 PThreads[15988:203200] ns timer in the thread: 4867
2018-04-14 11:02:51.353472+0800 PThreads[15988:203200] perform block in thread: 4867
2018-04-14 11:02:51.839917+0800 PThreads[15988:203206] gcd timer in the thread: 9219

虽然并行队列的async功能并不会因为启动了runloop受到影响，但是可以发现如果不去保存runloop，这个保活的线程除了定时器能正常处理之外，其他时候不会再被GCD复用

使用port保活

如果不使用NSTimer这种稳定的唤醒机制来保活线程，而是采用port的方式，线程的表现是否依旧符合预期？

__block CFRunLoopRef serialRunLoop = NULL;
dispatch_queue_t mainQueue = dispatch_get_main_queue();
dispatch_queue_t serialQueue = dispatch_queue_create("concurrent.queue", DISPATCH_QUEUE_CONCURRENT);

dispatch_async(serialQueue, ^{
    serialRunLoop = [NSRunLoop currentRunLoop].getCFRunLoop;
    [[NSRunLoop currentRunLoop] addPort: [NSPort new] forMode: NSDefaultRunLoopMode];
    [[NSRunLoop currentRunLoop] runUntilDate: [NSDate dateWithTimeIntervalSinceNow: 600]];
});

dispatch_source_t timer = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, mainQueue);
dispatch_source_set_timer(timer, DISPATCH_TIME_NOW, 0.5 * NSEC_PER_SEC, 0 * NSEC_PER_SEC);
dispatch_source_set_event_handler(timer, ^{
    CFRunLoopPerformBlock(serialRunLoop, NSDefaultRunLoopMode, ^{
        NSLog(@"perform block in thread: %d", mach_thread_self());
    });
});
dispatch_resume(timer);

此时任务总是不会被处理。由于runloop需要被唤醒才能处理队列任务，而perform block只是单纯的添加任务，没有唤醒功能。为了线程能够继续执行任务，这时候还需要不断的wake up线程：

dispatch_source_t timer = dispatch_source_create(DISPATCH_SOURCE_TYPE_TIMER, 0, 0, mainQueue);
dispatch_source_set_timer(timer, DISPATCH_TIME_NOW, 0.5 * NSEC_PER_SEC, 0 * NSEC_PER_SEC);
dispatch_source_set_event_handler(timer, ^{
    CFRunLoopPerformBlock(serialRunLoop, NSDefaultRunLoopMode, ^{
        NSLog(@"perform block in thread: %d", mach_thread_self());
    });
    CFRunLoopWakeUp(serialRunLoop);
});
dispatch_resume(timer);

/// 日志输出
2018-04-14 11:15:10.314064+0800 PThreads[19459:248764] perform block in thread: 2563
2018-04-14 11:15:10.763480+0800 PThreads[19459:248764] perform block in thread: 2563
2018-04-14 11:15:11.263651+0800 PThreads[19459:248764] perform block in thread: 2563
2018-04-14 11:15:11.763957+0800 PThreads[19459:248764] perform block in thread: 2563
2018-04-14 11:15:12.264006+0800 PThreads[19459:248764] perform block in thread: 2563
2018-04-14 11:15:12.763384+0800 PThreads[19459:248764] perform block in thread: 2563

结论

从测试来看，在子线程启动runloop并不是一个很明智的选择：这会导致线程保活期间被遗弃，失去了处理消息派发的能力，且无法响应其他线程的通信。其次，即便可以通过perform block来继续为保活线程添加任务处理，但在保活线程的runloop缺乏稳定的唤醒机制的情况下，还需要其他线程来提供唤醒能力，这增加了代码设计的成本，并且不会有额外的好处