Fix race in ThreadSafeSingleton

[rda-dev]

The self.__storage attribute may be overwritten by the ThreadSafeSingleton.reset() method and thus will return a None from the _provide(...) method.

Additionally, we should place a return statement inside the lock to prevent reading the shared variable. A more efficient solution is 'double-checked locking optimization.'

[coveralls]

Coverage increased (+0.03%) to 89.906% when pulling f28abb2 on rda-dev:fix-thread-safe-singleton into 4120afd on ets-labs:develop.

[rmk135]

Hey @rda-dev ,

Thanks for the PR. This looks very good. A cosmetic question before merging - do we need storage temp variable? Does it have any special meaning?

If it's not required, can you, please, remove it in favour of ```self.__storage``?

[rda-dev]

Hi @rmk135 ,

Yes, the first storage local variable assignment captures an attribute value and allows not to use lock if the variable is already initialized. The second storage assignment may be replaced with return self.__storage statement, but I think it will be clearer to leave a single return from the method and reuse the storage variable here. Isn't it?

[rmk135]

@rda-dev I like to avoid temporary variables when possible. Here is what I meant:

 if self.__storage is None: with self.__storage_lock: if self.__storage is None: self.__storage = __factory_call(self.__instantiator, args, kwargs) return self.__storage

It will work the same way, right?

[rda-dev]

No, it's not the same. In your example, when you release the lock, the parallel thread may set the __storage attribute to None (with reset() method) and you will return None instead of an initialized object.

[rmk135]

Heh, that's tricky. I got the difference finally. Didn't consider reset() effect eventually. Thanks for explanation. Merging.

[rmk135]

I will cythonize the code and make a release in a couple of hours.

[rmk135]

@rda-dev I have released fix in version 4.3.7. See changelog https://python-dependency-injector.ets-labs.org/main/changelog.html.

I have also added you to the list of contributors https://github.com/ets-labs/python-dependency-injector/blob/master/CONTRIBUTORS.rst. It is distributed with every copy of Dependency Injector.

Thank you again for the fix and for explaining how it works.

Best,
Roman