Recursive / Iterative NSURLSessionDataTask causing memory leak

Question

I am having a problem with memory leaking in my code, I have a need to GET many URL's in quick succession, each GET is influenced by the result of the previous GET. The purpose is to look for a specific piece of content within the response.

I found the cleanest way to implement this is recursively, as I can use the same method to identify if the desired value is present in the response. Functionally it works very well, but it leaks memory as described below. I have also implemented the same functionality in an iterative fashion, and this also leaks memory.

To my mind it seems that the NSURLSession API is responsible for leaking this memory, and it only occurs when multiple calls are made in very quick succession. However, I would appreciate if anyone can point out any obvious mistakes I am making.

Update 10/09/14:

Updated to add a recursion counter, demonstrating the leak still occurs even if the code isn't executed an infinite number of times. Also tidied up the implementation slightly, re-using the NSURLSession and NSURLSessionConfiguration as properties within the view controller.

Sample Code:

- (void)performURLCallRecursive {

    recursionLimiter++;

    if (recursionLimiter > 10) {
        [self.session finishTasksAndInvalidate];
        return;
    }

    NSURL * checkURL = [NSURL URLWithString:@"http://www.google.com"];

    __block NSMutableURLRequest * urlRequest = [[NSMutableURLRequest alloc] initWithURL:checkURL
                                                                            cachePolicy:NSURLRequestReloadIgnoringLocalCacheData
                                                                        timeoutInterval:0.0f];

    __weak typeof(self) weakSelf = self;

    NSURLSessionDataTask * task = [self.session dataTaskWithRequest:urlRequest
                                                  completionHandler:^(NSData *data, NSURLResponse *response, NSError
*error) {

                                                      NSString * body = [[NSString alloc] initWithData:data encoding:NSUTF8StringEncoding];
                                                      NSLog(@"Body: %@", body);

                                                      [weakSelf performURLCallRecursive];

                                                  }];

    [task resume];
 }

#pragma mark - Getters

- (NSURLSessionConfiguration *)sessionConfiguration {

    if (!_sessionConfiguration) {

        _sessionConfiguration = [NSURLSessionConfiguration ephemeralSessionConfiguration];

        [_sessionConfiguration setAllowsCellularAccess:NO];
        [_sessionConfiguration setTimeoutIntervalForRequest:10.0f];
        [_sessionConfiguration setTimeoutIntervalForResource:10.0f];

        [_sessionConfiguration setURLCache:[[NSURLCache alloc] initWithMemoryCapacity:0 diskCapacity:0 diskPath:nil]];

    }

    return _sessionConfiguration;
 }

- (NSURLSession *)session {

    if (_session == nil) {

        _session = [NSURLSession sessionWithConfiguration:self.sessionConfiguration
                                                 delegate:[SPRSessionDelegate new]
                                            delegateQueue:nil];

    }

    return _session; 
}

The memory leaks as reported by instruments. (NB: These vary slightly every time, but for the most part contain the same leaks, just more or less of the same leaks):

Further Update:

So, I actually implemented the same code iteratively, and the memory leak still occurs. For this example I included a loop limiter so it doesn't execute for ever. Can anyone help me figure out what on earth is going on here?

- (void)performURLCallIterative
{

    int loopLimiter = 0;

    do {

        NSURLSessionConfiguration * defaultSession = [NSURLSessionConfiguration defaultSessionConfiguration];

        [defaultSession setAllowsCellularAccess:NO];
        [defaultSession setTimeoutIntervalForRequest:10.0f];
        [defaultSession setTimeoutIntervalForResource:10.0f];

        NSURLSession * session = [NSURLSession sessionWithConfiguration:defaultSession
                                                               delegate:self
                                                          delegateQueue:nil];


        NSURL * checkURL = [NSURL URLWithString:@"http://google.com"];

        NSMutableURLRequest * urlRequest = [[NSMutableURLRequest alloc] initWithURL:checkURL
                                                                        cachePolicy:NSURLRequestReloadIgnoringLocalCacheData
                                                                    timeoutInterval:0.0f];

        __weak NSURLSession * weakSession = session;

        dispatch_semaphore_t semaphore = dispatch_semaphore_create(0);

        NSURLSessionDataTask * task = [session dataTaskWithRequest:urlRequest
                                                 completionHandler:^(NSData *data, NSURLResponse *response, NSError *error) {

                                                     NSString * body = [[NSString alloc] initWithData:data encoding:NSUTF8StringEncoding];
                                                     NSLog(@"Body: %@", body);

                                                     dispatch_semaphore_signal(semaphore);

                                                     [weakSession invalidateAndCancel];

                                                 }];

        [task resume];

        dispatch_semaphore_wait(semaphore, DISPATCH_TIME_FOREVER);

        loopLimiter++;

    } while (loopLimiter <= 6);

}

Update 10/09/14:

This is still occurring on iOS 8 for any Googlers who may have found their way here. As far as I am concerned this is a bug in iOS.

Answer 1

I had a lot of problems with memory from NSURLSession and I finally fixed it by not using a new session for each request. Sessions are generally defined on Wikipedia as:

a semi-permanent interactive information interchange

As such, Apple's convenience class method [NSURLSession sharedSession] gives us a clue of how NSURLSession objects are intended to be used: as semi-permanent objects, not one-off objects created fresh for each request, like you are doing.

You are making a new session object per request for a ton of requests that, from the server's perspective, are all part of a single session with a single client.

I was doing the same thing until I realized this was the source of my woes. I did not find Apple's documentation on this very clear, but after I realized the error of my ways, it made certain things in the documentation suddenly make more sense, like why there is a sharedSession singleton convenience method of NSURLSession, why the word "tasks" is plural in finishTasksAndInvalidate, why they called it a "session", why it has a cache, etc. (If it was just for one request, why would it be a "session" and what good would a "cache" be?)

It helps to know how a browser like Safari looks at a session. A new session starts the first time you make a connection to a given server. Setting up the session involves creating a cache of SSL certificates, establishing authentication, handshaking, etc. It would be extraordinarily inefficient to do all of that every time some JavaScript on a page makes a new request to the same server, especially since modern web apps constantly make requests with callbacks etc. That is why a single session is established for a whole huge set of requests and responses -- a conversation, if you will, between the client and server. Eventually, a session expires, but usually this happens after several minutes, not after one request!

The point is, how you should be using NSURLSession objects is to make a singleton with a strongly referenced NSURLSession object as a property. Do this if you need to customize the session's configuration, (like turning caching off, etc.). However if you do not need to customize it, just use Apple's sharedSession.

If you use a singleton on a custom class, then, if you never need to set the session property to nil, then you never need to invalidateAndCancel or finishTasksAndInvalidate. Instead, just resetWithCompletionBlock or flushWithCompletionBlock to clear out connection caches periodically.

If you hate singletons you can still use a session as a property, just make sure to invalidateAndCancel or finishTasksAndInvalidate the session before its last owner gets deallocated by the ARC runtime.

Also note that setting your NSURLSession object's URLCache property to nil is the proper way to shut off caching. That's what Apple says they do for backgroundSessionConfiguration.

See my other answers on this topic here and here.

Answer 2

A developer support request to Apple reveals this to be a bug within iOS 7. There is no fault with the code sample posted above (Either recursively or iteratively) and it has reportedly been fixed in the iOS 8 GM release.

Update:

This is still occurring in iOS 8.1

Answer 3

- Update 9/12/2014

Solution: wait for iOS8.

- Update 9/10/2014

Whoa, this is spiraling into some Nth dimension of complexity :P. I hope one way or another you get a break here quick.

I have a few other things for you to try.

1) Could you make sure NSZombies is turned off. In Xcode, Product->Scheme->Edit Scheme...->Enable Zombie Objects (NOT ticked).

2) Also try cachePolicy:NSURLCacheStorageNotAllowed for your NSMutableURLRequest.

3) Could you see if you are completing with an error? Just put this around your body string assignment...

if (error == nil)
{
    //Enter data->string code here
}

4) Could you see if you are not getting status 200?

NSInteger statusCode = [(NSHTTPURLResponse *)response statusCode];

5) It is hard to picture exactly how your project is set up. I would have an NSObject type class that houses the NSURLSession methods, which is separate from the UIViewController class from which it is being called. The timer or whatever recursion method you wish to choose would then call the url session associated methods from the UIViewController.

- Update 9/9/2014

You are correct about my question (2). The data task is resumed before completion and after the data task completes the session is invalidated. I haven't seen it done this way, but it makes sense. Just tested on my end, no leaks with regards to [session invalidateAndCancel]...

Could you check that your completion handler executes? Perhaps it doesn't and the session is never cancelled before a new task is started?

I am noticing that there are a few references to HTTP Headers in the Instruments Leaks report, maybe if you are not specifying either a [urlRequest setHTTPMethod:@"GET"] the request is missing some basic headers?

(I'll edit after we find the solution, so this doesn't look like a discussion).

- Original 9/8/2014

Interesting question! I have troubleshot leaks associated with NSURLSessions. Definitely @autoreleasepool{} and others are good suggestions to try so far... But!

I am afraid the thing you asked us to look past might be the culprit here.

Just a few observations first:

1) It is not clear to me why you would need to __weak the self here. What is the retain cycle you are trying to avoid? Perhaps this is more clear in the code you are actually using aside from your "sample".

2) What is the reason for the call to invalidate the session before the data task associated with that session even has a chance to complete, let alone resume. The data task is in the suspended state until resumed.

3) If you are recursively running a method like this, then I think it is crucial to specify or at least consider what delegate queue, otherwise having it set to nil defaults it to serial operation queue. What happens when the delegate calls before the completion handler finishes, in an infinite loop - most likely a huge pile up.

--

I believe that the main issue here is that you are starting a new or canceling the NSURLSessionDataTask before it has a chance to complete. Look at +sesssionWithConfiguration:

(sorry can't include pictures yet, hopefully after this answer)

https://developer.apple.com/library/ios/documentation/Foundation/Reference/NSURLSession_class/Introduction/Introduction.html#//apple_ref/occ/clm/NSURLSession/sessionWithConfiguration:

The point is here...

Important

The session object keeps a strong reference to the delegate until your app explicitly invalidates the session. If you do not invalidate the session by calling the invalidateAndCancel or resetWithCompletionHandler: method, your app leaks memory.

My suggestion to try is...

 //Your code above...
    [task resume];
    [session finishTasksAndInvalidate];
}

In theory this should prevent any new sessions from starting before completion, according to the description, "...new tasks cannot be created in the session, but existing tasks continue until completion. After the last task finishes and the session makes the last delegate call, references to the delegate and callback objects are broken..."

I am still not sure about invalidating the session before resuming it.

I hope this helps. Good luck.

Answer 4

The only suggestions I have are perhaps using an @autoreleasepool{} and converting the __weak id self to __block id self. I don't think the __block vs. __weak will do anything differently, but give it a shot.

I'm not sure what one should expect with ARC when running something asynchronously AND recursively. Looking at other questions with asynchronous recursive calls and ARC, there isn't any consistent solution. Take a look here, for example.