I believe this was related to #2638.

I'll try and dig in today or tomorrow.

It looks like @mattt actually removed the dispatch_once in commit ace91df

I'm guessing we need still a dispatch_once, just inside the conditional. Something like this:

+ (void)initialize {
    if ([NSURLSessionTask class]) {
        static dispatch_once_t onceToken;
        dispatch_once(&onceToken, ^{
            NSURLSessionDataTask *dataTask = [[NSURLSession sessionWithConfiguration:nil] dataTaskWithURL:nil];
            Class taskClass = [dataTask superclass];

            af_addMethod(taskClass, @selector(af_resume),  class_getInstanceMethod(self, @selector(af_resume)));
            af_addMethod(taskClass, @selector(af_suspend), class_getInstanceMethod(self, @selector(af_suspend)));
            af_swizzleSelector(taskClass, @selector(resume), @selector(af_resume));
            af_swizzleSelector(taskClass, @selector(suspend), @selector(af_suspend));

            [dataTask cancel];
        });
    }
}

Does this resolve the issue for you?

It actually may be more complicated than that... Let me do some more thinking.

After further review, I THINK the dispatch_once will solve your issue as well as the other linked issues here, but I need verification.

Could you take a look at this branch, or drop in this patch and verify things look good on your end?

Your patch is deadlocking for me @kcharwood in the AFNetworking iOS Example app.

Deadlock Case

If you use this modified patch, and run the example app, I'm assuming you will see the same behavior.

+ (void)initialize {
    NSLog(@"Attempting to initialize: %@", NSStringFromClass([self class]));

    static dispatch_once_t onceToken;
    dispatch_once(&onceToken, ^{
        NSLog(@"Initializing class: %@", NSStringFromClass([self class]));

        NSURLSessionDataTask *dataTask = [[NSURLSession sessionWithConfiguration:nil] dataTaskWithURL:nil];
        Class taskClass = [dataTask superclass];

        af_addMethod(taskClass, @selector(af_resume),  class_getInstanceMethod(self, @selector(af_resume)));
        af_addMethod(taskClass, @selector(af_suspend), class_getInstanceMethod(self, @selector(af_suspend)));
        af_swizzleSelector(taskClass, @selector(resume), @selector(af_resume));
        af_swizzleSelector(taskClass, @selector(suspend), @selector(af_suspend));

        [dataTask cancel];
    });
}

Produces the following output and hangs:

2015-05-07 09:24:10.692 AFNetworking iOS Example[50814:37472160] Attempting to initialize: NSURLSessionTask
2015-05-07 09:24:10.692 AFNetworking iOS Example[50814:37472160] Initializing class: NSURLSessionTask
2015-05-07 09:24:10.692 AFNetworking iOS Example[50814:37472160] Attempting to initialize: __NSCFLocalSessionTask

Static State Variable

One way to avoid the deadlock would be to store a second static variable to early out if necessary.

+ (void)initialize {
    NSLog(@"Attempting to initialize: %@", NSStringFromClass([self class]));

    static BOOL isInitialized = NO;

    if (isInitialized) {
        return;
    }

    static dispatch_once_t onceToken;
    dispatch_once(&onceToken, ^{
        isInitialized = YES;

        NSLog(@"Initializing class: %@", NSStringFromClass([self class]));

        NSURLSessionDataTask *dataTask = [[NSURLSession sessionWithConfiguration:nil] dataTaskWithURL:nil];
        Class taskClass = [dataTask superclass];

        af_addMethod(taskClass, @selector(af_resume),  class_getInstanceMethod(self, @selector(af_resume)));
        af_addMethod(taskClass, @selector(af_suspend), class_getInstanceMethod(self, @selector(af_suspend)));
        af_swizzleSelector(taskClass, @selector(resume), @selector(af_resume));
        af_swizzleSelector(taskClass, @selector(suspend), @selector(af_suspend));

        [dataTask cancel];
    });
}

Produces the following output and runs properly:

2015-05-07 09:26:12.725 AFNetworking iOS Example[50861:37477359] Attempting to initialize: NSURLSessionTask
2015-05-07 09:26:12.725 AFNetworking iOS Example[50861:37477359] Initializing class: NSURLSessionTask
2015-05-07 09:26:12.725 AFNetworking iOS Example[50861:37477359] Attempting to initialize: __NSCFLocalSessionTask
2015-05-07 09:26:12.725 AFNetworking iOS Example[50861:37477359] Attempting to initialize: __NSCFLocalDataTask
2015-05-07 09:26:12.728 AFNetworking iOS Example[50861:37477413] Attempting to initialize: __NSCFLocalUploadTask
2015-05-07 09:26:12.728 AFNetworking iOS Example[50861:37477413] Attempting to initialize: NSURLSessionDataTask

Another option would be to compare the class against NSURLSessionTask.

Class Check

+ (void)initialize {
    NSLog(@"Attempting to initialize: %@", NSStringFromClass([self class]));

    if ([NSStringFromClass([self class]) isEqualToString:NSStringFromClass([NSURLSessionTask class])]) {
        NSLog(@"Initializing class: %@", NSStringFromClass([self class]));

        NSURLSessionDataTask *dataTask = [[NSURLSession sessionWithConfiguration:nil] dataTaskWithURL:nil];
        Class taskClass = [dataTask superclass];

        af_addMethod(taskClass, @selector(af_resume),  class_getInstanceMethod(self, @selector(af_resume)));
        af_addMethod(taskClass, @selector(af_suspend), class_getInstanceMethod(self, @selector(af_suspend)));
        af_swizzleSelector(taskClass, @selector(resume), @selector(af_resume));
        af_swizzleSelector(taskClass, @selector(suspend), @selector(af_suspend));

        [dataTask cancel];
    }
}

Produces the following output and runs properly:

2015-05-07 09:27:16.067 AFNetworking iOS Example[50895:37480604] Attempting to initialize: NSURLSessionTask
2015-05-07 09:27:16.067 AFNetworking iOS Example[50895:37480604] Initializing class: NSURLSessionTask
2015-05-07 09:27:16.067 AFNetworking iOS Example[50895:37480604] Attempting to initialize: __NSCFLocalSessionTask
2015-05-07 09:27:16.067 AFNetworking iOS Example[50895:37480604] Attempting to initialize: __NSCFLocalDataTask
2015-05-07 09:27:16.068 AFNetworking iOS Example[50895:37480655] Attempting to initialize: __NSCFLocalUploadTask
2015-05-07 09:27:16.068 AFNetworking iOS Example[50895:37480655] Attempting to initialize: NSURLSessionDataTask

There's probably a better way to compare class equality, but that was the best solution I could come up with on short notice. See thread for more info.

Another option would be to switch from initialize to load, but that seems to have many other implications from past issues I've been digging through. They may have had other issues though due to the previous implementations of this workaround.

IMO, I think the [Class Check](#Class Check) approach is the best way to go.

Thoughts?

I think Class Check looks pretty good as well.

I put together a quick test to try and verify this in AFURLSessionManagerTests:

- (void)testSwizzlingIsWorkingAsExpected {
    [self expectationForNotification:@"com.alamofire.networking.task.suspend"
                              object:nil
                             handler:nil];
    NSURL *delayURL = [self.baseURL URLByAppendingPathComponent:@"delay/1"];
    NSURLSessionDataTask *task = [self.manager dataTaskWithRequest:[NSURLRequest requestWithURL:delayURL]
                                                 completionHandler:nil];
    [task resume];
    [task suspend];
    [self waitForExpectationsWithTimeout:2.0 handler:nil];
    [task cancel];


    [self expectationForNotification:@"com.alamofire.networking.task.suspend"
                              object:nil
                             handler:nil];

    NSURLSessionDataTask *uploadTask = [self.manager uploadTaskWithRequest:[NSURLRequest requestWithURL:delayURL]
                                                                   fromData:nil
                                                                   progress:nil
                                                          completionHandler:nil];
    [uploadTask resume];
    [uploadTask suspend];
    [self waitForExpectationsWithTimeout:2.0 handler:nil];
    [uploadTask cancel];
}

When running this test in isolation, this test fails on 2.5.3. When running with the class check patch, it passes.

I'm not sure this is a great test to add to the suite, because its still susceptible to race conditions from other tests. For example, if both of these classes had already been loaded from previous tests, and an odd number of session task classes had already been created, this test would pass on 2.5.3. Anyone have any better ideas on how to build a better test?

I updated the branch with @cnoon's patch for class name check.

I'm also wondering if the code can be reduced to

+ (void)initialize {
    if ([NSStringFromClass([self class]) isEqualToString:NSStringFromClass([NSURLSessionTask class])]) {

        af_addMethod([NSURLSessionTask class], @selector(af_resume),  class_getInstanceMethod(self, @selector(af_resume)));
        af_addMethod([NSURLSessionTask class], @selector(af_suspend), class_getInstanceMethod(self, @selector(af_suspend)));
        af_swizzleSelector([NSURLSessionTask class], @selector(resume), @selector(af_resume));
        af_swizzleSelector([NSURLSessionTask class], @selector(suspend), @selector(af_suspend));
    }
}

but I'm not super clear what the original need was to call "superclass" to get something that will work here. Still trying to parse through #2638

One more alternative a coworker suggested.

We could create a dummy class only used for swizzling. That way we don't have to make a protocol on NSURLSessionTask and swizzle in initialize, and we can swizzle from load in the dummy class only if NSURLSessionTask is available. I think that would address the original problems for the crashers, and would eliminate the multiple calls to initialize.

Something like this commit. I dropped a bunch of tests in there as well.

Oh man I like it! I think it could safely be condensed down to the following:

@implementation _AFURLSessionTaskSwizzling

+ (void)load {
    Class urlSessionTaskClass = [NSURLSessionTask class];

    af_addMethod(urlSessionTaskClass, @selector(af_resume),  class_getInstanceMethod(urlSessionTaskClass, @selector(af_resume)));
    af_addMethod(urlSessionTaskClass, @selector(af_suspend), class_getInstanceMethod(urlSessionTaskClass, @selector(af_suspend)));
    af_swizzleSelector(urlSessionTaskClass, @selector(resume), @selector(af_resume));
    af_swizzleSelector(urlSessionTaskClass, @selector(suspend), @selector(af_suspend));
}

@end

The load method will never get run more than once. Additionally, there's no need to add a guard around the Class object. Last tweak I made was to rename the variable to urlSessionTaskClass.

Thanks @cnoon. Just pushed an update with that refactor, along with one other note. On OS X, Apple says don't use the class method to check for existence, but rather NSClassFromString, so I made that update as well.

In OS X (and in iOS projects that do not meet the set of conditions just listed), you cannot use the class method to determine if a weakly linked class is available. Instead, use the NSClassFromString function in code similar to the following:

Class cls = NSClassFromString (@"NSRegularExpression");
if (cls) {
    // Create an instance of the class and use it.
} else {
    // Alternate code path to follow when the
    // class is not available.
}

Looks good @kcharwood! Thanks for being so thorough on this. Also, good to know on the class check 👍🏻

It looks like Travis just can't handle the XCTests with expectationsForNotifications. I've commented those out for now, since Travis is running Xcode 6.1.

The tests do work locally for me.

@fouvrai I think you may be right, as I am seeing some strange behavior on iOS 7. How many different ways can I possibly skin this cat...

I'm starting to question if this has every actually worked properly for iOS 7. I have an example that works properly for NSURLSessionDataTask, but trying to use UploadTask or DownloadTask causes a problem on iOS 7. It all works great on iOS 8.

Running various variations of the swizzling I've seen in the history of this file seems to fail in the same way. This one is nasty, mainly due to the implementation differences under the hood between iOS 8 and iOS 7 that Apple has hidden away from us.

I've built out a class diagram for NSURLSession for iOS 7 and iOS 8, based on what classes are being given to me at runtime.

iOS 7:

iOS 8:

Note the grey boxes, which represent the hierarchy if you do something like [[NSURLSessionDataTask class] superclass]. The green boxes represent the class hierarchy if you take an actual object returned by NSURLSessionManager, and go up the superclass tree. I have tests confirming the swizzling approach works for all green boxes.

Note the one red box in iOS 7 - The upload task object. Building out this diagram helped me identify that as the one failing test. If I disable the tests for UploadTask on iOS 7, everything else passes. If I run the UploadTask tests in iOS 7 in isolation or in the suite, it causes a deadlock inside of af_suspend.

I have not been able to fully diagnose why this is happening, but I'm using this ticket as a scratchpad to talk out my notes. I'm trying to determine if there is something different about the implementation of suspend in iOS 7 for the upload task that is causing a problem with the swizzle.

I know I might be commenting on this without the full history or context as to what's going on with this particular bit of code, but when someone says 'runtime', my ears perk up.

So to summarize, it appears that the entire purpose of the swizzled methods is to:

1. Add the af_resume/af_suspend methods to all (what we think is) the root class for all NSURLSessionTask objects used by the user
2. Call those newly added methods on the existing implementations of -[NSURLSessionTask resume] and -[NSURLSessionTask suspend] in order to post a notification

In addition, the reason that +initialize is being used as opposed to +load is due to not actually knowing what the true "root" class ("root" here being the closest superclass after NSObject) of the hierarchy tree is. This is due to the class hierarchy diagrams that Kevin drew up since +load is called in order of hierarchy, and this breaks on iOS 7.

Is there a reason we need to swizzle this into the class of NSURLSessionTask, so that the new functionality is to be used on all classes, even those that are created outside of AFNetworking? Could we simply add and swizzle in the same exact methods, but on a per-instance basis in the *TaskWithRequest methods in AFURLSessionManager, where everything is funneling into? We could then add/swizzle on the actual returned class (once) vs an unknown root class (which would run an unknown number of times in a runtime method like +initialize).

Bad? Depends on if this functionality must live on the furthest superclass and if all *SessionTask subclasses in the bundle need this functionality regardless of association with AFNetworking.

Another solution could be to actually use the runtime to find the furthest superclass that actually implements suspend/resume and add the af_* methods to just that one class. Then use the existence of those af_* methods we just added to know if we have already performed this action, or also wrap this action into a dispatch_once.

I'm also not positive what is causing the deadlocking referred to above. I'm assuming the onceToken, since it is static and only found in that one category, is actually the same token across multiple calls to +initialize. This would mean that +intialize is called concurrently?

Removing the dispatch_once is OK. The normal way that +initialize is protected so that the code is run only once is like this

+(void)initialize
{
if ([self class] == [NSURLSessionTask class] {
// code that should only run once goes here
}
}

I think in the code committed by @mattt ace91df he had a typo where the == [NSURLSessionTask class] part was missing.

This entire issue is the cause of my related issue: https://github.com/AFNetworking/AFNetworking/issues/2660 AFNetworkActivityIndicatorManager doesn't work for background downloads.

If you change this code again I request that you make sure it works for background downloads.

I think the swizzling can be made to work for background downloads if you swizzle [self class] one time.

My real opinion is that we should throw away the swizzling code and go back to the KVO code. KVO observing of the task state is the right way to do this. I understand there were some crashes in some cases when removing the observers. I expect that those crashes could be fixed. Swizzling is an attempt to avoid fixing the KVO code. Everyone at Apple says don't swizzle. They're obviously right.

Those crashes are not fixed—they're a bug in KVO itself...and that bug is still at large, as of iOS 8.1. In our experience, the KVO crashes increase our per-session crash rate by ~0.2% (leading to hundreds of crashes a day), making them our most common crash by a factor of 10.

I agree that swizzling is a terrible, terrible idea. But ultimately, code quality is less important than user experience, which is why swizzling is unfortunately the right thing to do here.

I pushed up a rebase off master with just the url session unit tests I have for this issue if anyone wants to try their solution on it.

https://github.com/AFNetworking/AFNetworking/tree/updated_url_session_tests

Running rake locally on machine shows all tests passing. Running on Travis shows failures. I can't catch a break... Looks like the failures are for 10.9.5. Trying to find a 10.9 machine now.

@phoney personally I'm not a huge fan of asking everyone to use custom resume/suspend methods, as that can be confusing and as you said, not backwards compatible. If we can "make it magically work", I'm more of a fan of that approach.

+1 @kcharwood. I think we all thought through alternative possibilities while maintaining backwards compatibility. That's why swizzling is really the only option since KVO is crashing. All other proposed solutions would have serious public API implications or would break this feature for all our current users.

TRAVIS CI JUST PASSED MY TESTS. I'm going to spend quite a bit of time cleaning this branch up now, and adding appropriate documentation.

Ok. Branch cleaned up. More tests added. Comments added to explain my thoughts.

I would love as many eyes as possible on this one.

I tested your newest code in my app on iOS 8 and iOS 7 and it seems to work fine. The network activity indicator spins for background downloads and non-background requests.

These changes all look good to me @kcharwood. I went through it with a fine tooth comb and it looks great! All the documentation you've added is really helpful, and the test coverage is really well done. Awesome job!

Looks good!

I'd like to thank everyone who jumped in on this issue and provided feedback/suggestions. Super nasty issue, and couldn't have gotten to this solution without everyones involvement.

cheers 🍻

af3.1.0version iOS10.2 still got the crash!