|Microsoft DirectX 9.0|
Any DirectShow application contains at least two important threads: the application thread, and one or more streaming threads. Samples are delivered on the streaming threads, and state changes happen on the application thread. The main streaming thread is created by a source or parser filter. Other filters might create worker threads that deliver samples, and these are considered streaming threads as well.
Some methods are called on the application thread, while others are called on a streaming thread. For example:
Having a separate streaming thread allows data to flow through the graph while the application thread waits for user input. The danger of multiple threads, however, is that a filter may create resources when it pauses (on the application thread), use them inside a streaming method, and destroy them when it stops (also on the application thread). If you are not careful, the streaming thread might try to use the resources after they are destroyed. The solution is to protect resources using critical sections, and synchronize streaming methods with state changes.
A filter needs one critical section to protect the filter state. The CBaseFilter class has a member variable for this critical section, CBaseFilter::m_pLock. This critical section is called the filter lock. Also, each input pin needs a critical section to protect resources used by the streaming thread. These critical sections are called streaming locks; you must declare them in your derived pin class. It is easiest to use the CCritSec class, which wraps a Windows CRITICAL_SECTION object and can be locked using the CAutoLock class. The CCritSec class also provides some useful debugging functions. For more information, see Critical Section Debugging Functions.
When a filter stops or flushes, it must synchronize the application thread with the streaming thread. To avoid deadlocking, it must first unblock the streaming thread, which might be blocked for several reasons:
Therefore, when the filter stops or flushes, it must do the following:
Flushing and stopping both happen on the application thread. A filter stops in response to the IMediaControl::Stop method. The Filter Graph Manager issues the stop command in upstream order, starting from the renderers and working backward to the source filters. The stop command happens completely inside the filter's CBaseFilter::Stop method. When the method returns, the filter should be in a stopped state.
Flushing typically occurs because of a seek command. A flush command starts from the source or parser filter, and travels downstream. Flushing happens in two stages: The IPin::BeginFlush method informs a filter to discard all pending and incoming data; the IPin::EndFlush method signals the filter to accept data again. Flushing requires two stages because the BeginFlush call is on the application thread, during which the streaming thread continues to deliver data. Therefore, some samples may arrive after the BeginFlush call. The filter should discard these. Any samples that arrive after the EndFlush call are guaranteed to be new, and should be delivered.
The sections that follow contain code samples showing how to implement the most important filter methods, such as Pause, Receive, and so forth, in ways that avoid deadlocks and race conditions. Every filter has different requirements, however, so you will need to adapt these examples to your particular filter.
Note The CTransformFilter and CTransInPlaceFilter base classes handle many of the issues described in this article. If you are writing a transform filter, and your filter does not wait on events inside a streaming method, or hold onto samples outside of Receive, then these base classes should be sufficient.