Code review comment for lp:~raof/mir/no-ipc-on-compositor-threads

OK, I've made some progress:

multi_stream is hitting the problem because it repeatedly creates new BufferStreams and destroying the old ones (and RenderSurfaces).

When BufferStreams are created and destroyed like this the buffer_cache in SessionMediator grows intermittently but inexorably.

Up until the point when file handles are exhausted closing the connection releases the cache and the buffers. (Once file handles are exhausted the server struggles to release the connection.)

I've not yet got my head around where the buffers created for mir_buffer_stream_get_graphics_region() are being tracked and should be released more consistently.