5.2 Reliability Assurances

A few basic rules keep the main server thread and the multi-rendering threads from deadlocking or colliding with each other.

• Semaphored memory allocation functions - these keep collisions from occurring in dynamic memory allocation and deallocation routines.

• One thread per client - no matter how many rendering contexts (OpenGL contexts, or PEX renderers) a client has, this ensures sequentiality of requests by only processing one protocol request at a time for each client. Therefore only one rendering thread is needed per client.

• Specialization - in order to minimize the number of locks (and opportunities for deadlock), some tasks are restricted to the main server thread. These include anything that manipulates the server's data structures. For example, any creating, destroying, and looking up X resources is only done by the main X server thread. OpenGL and PEX rendering requests are likewise only executed by the client's rendering thread.

  This creates certain implementation requirements for extension request processing functions. For example, in the OpenGL extension the main X server thread decodes protocol requests, looks up X resource IDs, and either increments reference counts or copies data from X internal structures so that a structure will never disappear while it is being referenced in a rendering thread. Even when a client connection closes, some data structures cannot be freed immediately. A special interface is provided so that the rendering thread can perform clean-up operations after the client connection is closed; only after the rendering thread clean-up action is done can the main X server thread be allowed to free the client's resources.

• Hand-off and rendezvous - the mechanisms described in the previous section allow the main thread to relinquish control of the client's shared data until it is informed that the rendering thread has completed its task. This cooperative processing eliminates the need for more locks.

• Pure procedures - rendering state data is unique per client. The only static variables allowed are those containing constant data. Any exception to this rule would require a lock, thus defeating potential concurrency among rendering threads.