[linux-audio-dev] Processing a chain of machines

Hi all, I'm currently working on creating Mess, a Buzz-like software studio written in C++ on Linux, and have run into a somewhat critical problem (note: this is my first "real" audio app): I'm using PortAudio as audio API and it's outputbuffer keeps underflowing because my callback function is too slow. One of my goals is to enable the user to make a network of synths, samples, effects,etc. ("machines" in Buzz-terminology). My initial idea for implementing this was to associate a buffer and a process-function with each machine. If the machine only generated sound, the process-function would simply fill the buffer, if it was an effect the buffer would contain the effect's input which would be consequently overwritten by the effect's output. The MachineManager class contained a processChain-function which was responsible for calling each machine's process-function in the right order and copying and mixing the contents of the buffers according to the connections in the network. PortAudio's callback called the processChain-function and next copied the contents of the master machine's buffer (the master represents the output in the network) to PortAudio's outputbuffer. As a test I made a simple sine tone generating machine and connected it to the master. This resulted in some weird noises when using small buffers and in clear sine tones with pauses inbetween when using very large buffers. In an attempt to let the callback do less work, I made an extra thread together with a set of two buffers with read/write flags, the idea being that the thread continuously tried to fill the buffers with the write flag set, using processChain (after this the buffer's flag would be set to 'read' ; if there was nothing left to write the thread would go to sleep) and the callback function given to PortAudio would check each time it ran for a readable buffer and copy that to it's own outputbuffer (and setting the just read buffer's flag to 'write', awakening the thread with processChain in it). This again resulted in some weird noises (mostly due to thread scheduling not quite going the way I expected). The problem with my first method seems to be the copying of all the buffers that makes the callback function too slow, but I don't immediately see any other way of doing it (is there one?). In retrospect the second way just seems wrong. So my question right now is: how can I process a chain of machines in a fast enough way? Raf