Revert "Fix for a race condition in URLSession (#949)"

[weissi]

This reverts commit 015ded4.

The reason to revert this is that it this pretty certainly works around another bug (SR-5759) and IMHO introduces two new bugs instead.

Dispatch Sources should be cancellable independently. If you have two DispatchSources watching for events on one and the same PR, the should be independent. That was seemingly the problem that was seen here, see the original bug and the discussion in #949.

The two new bugs introduced:

• a blocking wait self.cancelInProgress?.wait() is introduced when creating a _MultiHandle. Creating a large number of _MultiHandles concurrently could lead to a lot of threads being blocked in the wait. That might exhaust libdispatch's thread pool which might then lead to the cancellation handlers never been fully run
• CFURLSessionMultiHandleAssign is run on the wrong thread (commandQueue.async )

this should only be merged when SR-5759 is done.

[pushkarnk]

Looks like swiftlang/swift-corelibs-libdispatch#295 was merged. Let me edit the title first

[pushkarnk]

I ran the test related to SR-4567 a couple of times and it does pass.

[pushkarnk]

@MadCoder I took a look at swiftlang/swift-corelibs-libdispatch#295 and have a fundamental question in the context of #949

A DispatchSource.cancel() call returns immediately, the actual EPOLL_CTL_DEL happens asynchronously, and the cancel handler invocation indicates the completion of the cancel operation. The problem in #949 was that after the cancel returned but before the EPOLL_CTL_DEL was executed, if we get EPOLL_CTL_ADD or EPOLL_CTL_MOD requests on the same fd, the latter are missed and we see a hang. Does your PR address this problem?

I did run the test case for #949 only with your fix and it goes through, but just wanted to be sure the problem was actually addressed.

[MadCoder]

Yes, I think my fix addresses this problem

[pushkarnk]

Thanks @weissi for this great piece of work. As you proved, #949 was vulnerable, because, IMO, fixing a race condition with added synchronisation often creates more race conditions :) I'll go ahead commit this revert.

[pushkarnk]

@swift-ci please test and merge