Re: [linux-audio-dev] Re: lock-free data structures

Juhana Sadeharju wrote: We use a very simple approach in LinuxSampler: in the disk thread if all ringbuffers are filled or no voice is active then we sleep for 30msec. It could be done with a semaphore too (or waiting for a byte from a FIFO) but it's simply not worth the trouble changing the current approach. When the disk streamer gets busy the usleep() is not called and if it is idle then periodical sleeps of 30msec aren't going stressing the process scheduler in any way. I know sleeping is not the "elegantest" method but it is the simplest one and since we plan to support multiple disk threads (one per physical disk) adding all the needed IPC in the audio thread to communicate with the disk threads (and since the datarate of the disk streams is dynamic (user moves the pitchbender etc) the disk thread has no easy way to exactly figure out when to wake up other than looking at the fill amount of each active ringbuffer. Just open the file engines/gig/DiskThread.cpp and search for: IsIdle (a boolean variable that indicates if the diskthread needs to sleep a bit or if it did perform disk I/O during the current cycle). The idling/sleeping mechanism consist of 3 lines of code, not that complex I'd say, and it works very well :) cheers, Benno http://www.linuxsampler.org