A missing Py_DECREF on the future's "done" method
caused both the bound method and the underlying
self instance of the future to be leaked when
awaiting non-panda futures (such as _asyncio.Future).
This change includes a simple new test addition
to catch this in the future.
This adds persistent wrapper support (introduced by the previous commit) to PythonTask, which makes it possible for reference cycles involving tasks to be found and destroyed.
The major caveat is that it always creates a reference cycle. This can be broken automatically if there is no more Python reference to it by the time the last C++ reference is dropped, but the other way around requires the garbage collector.
For tasks, I think this it is generally the case that the last reference is in C++, since tasks are usually created and then handed off to the C++ task manager, and for applications that don't want to rely on the GC, it is easy to work around. If this turns out to be a problem, though, we can add a special garbage collection pass to the task manager.
This change enables persistent Python wrapper objects for Python subclasses of typed, reference counted C++ objects. That means that these objects will store a reference to `self` on the C++ object, and interrogate will always return that instead of making a new Python wrapper every time it is returned from C++.
Practically, this means that you could subclass eg. PandaNode, Event, Fog, what have you, and store these in the scene graph - the Python data in the subclass will be retained and Panda will return your subclass when you ask for the object back rather than creating a new wrapper object without your original data.
To do this, Interrogate generates a proxy class inheriting from the C++ type with additional room to store a `self` pointer and a TypeHandle (which is returned by an overridden `get_type()`). This TypeHandle is automatically created by registering the Python subclass with the typing system. The proxy class is only used when the constructor detects that it's constructing for a subtype, so that regular uses of the C++ type are not affected by this mechanism.
(The registration with the typing system could use some improvement. There's no regard for namespacing right now, in particular. Furthermore, we could move the registration to an `__init_subclass__()` method, with parameters to specify an existing TypeHandle or to customize the type name.)
This creates a reference cycle, which must be cleared somehow. One way this happens is by overriding `unref()` in the proxy, which checks that if the Python and C++ reference counts are both 1, this must be the circular reference, and then it breaks the cycle. Note that this will _only_ work if all other Python references have been cleared before the last C++ reference goes away. For the other case, we need to rely on Python's garbage collector, so these classes also implement tp_traverse and tp_clear.
This commit also therefore re-enables Python garbage collector support (ie. defining `__traverse__()`), which was previously disabled due to the problems caused by multiple Python wrappers referring to the same C++ object. To avoid these problems, Panda will only cooperate with Python's GC if the C++ reference count is 1, in which case we can trivially prove that the last reference must be from the Python wrapper. Note that this explains why we need persistent wrappers for traversal to work--if there are multiple Python wrappers pointing to the C++ object, and they are all participating in a reference cycle, the reference count cannot be 1, and cycle detection does not work. By default, the GC is only enabled for Python subclasses, but a class may define `__traverse__()` in order to opt-in, keeping the previous limitation in mind.
There is a second mechanism introduced by this commit: classes may define a public `__self__` member of type `PyObject *` if they want to use persistent objects even if they aren't being subclassed. This allows these classes to participate in Python's GC and avoid the situation above.
The cycle detection is implemented for PandaNode's Python tags, but this is very incomplete, since the traversal doesn't recurse into children and it won't work if there is more than one wrapper object that is part of a cycle, since PandaNode by default doesn't use persistent wrappers. This problem may need more attention later.
Fixes#1410
* Passing a Python subclass of a C++ class now works, the extra Python stuff isn't just discarded
* EventParameter objects are no longer automagically unwrapped - there's no more reason to pass an EventParameter to this method anyway, and it might be unexpected if it is treated specially.
This introduces AsyncFuture as a new base class of AsyncTask. It's modelled after asyncio's Future class, except that it is thread-safe and you can use result() to block the current thread waiting for the future to finish (of course this is not necessary for use with coroutines).
AsyncFuture should be used for any operation that finishes in the future, to get the benefit of awaitability within coroutines as well as a standard interface for querying status and results of the operation as well as cancelling it. As such, it's been implemented in various places, including texture.prepare() and win.trigger_copy().
Note that AsyncFuture is intended to be used *once*; it cannot be used more than once. As an example of how this works, tex.prepare() will return the same future as long as the prepare isn't complete, but when it is done, subsequent calls to tex.prepare() will return a new future.
rdb
John C. Allwein