>> The apps already need to do some type of synchronization internally. >> For example a player's disk thread, when its ringbuffer is full, needs >> to wait for the process thread to consume some data and thus free up > > Depends. If both ends are periodic processes no other synchronisation > is required. And e.g. Jack callback is such a process, and likely to > be one end.
How about the other "end" (i.e. the "disk thread"?) Would that normally be periodic? OK, even if your disk thread is periodic for some reason, how does that argue for library-level synchronization, *instead of* app-level synchronization? In this case the cost would be the same -- no loss. > You may be right about the (HW as opposed to compiler) re-ordering of > data w.r.t. pointers on some architectures. But AFAIK, at least on Intel > and AMD writes are not re-ordered w.r.t. other writes from the same CPU, "From the same CPU"? Are we regressing to non-SMP-only schemes? And "Intel and AMD" only? How about multiple cores / CPUs / caches? Pipeline reordering is not the main concern (though it can happen) -- cache coherence is. > Regarding the volatile declarations, at least on my version (which > is slightly different from Jack's) there is no performance penalty. Under which access patterns, with what compiler / optimization flags etc? I would not make such generalizations... Volatile frustrates the optimizer's ability to chose the optimal access patterns. > So I keep them just as reminders that these data are shared and may > change in unexpected ways. Hijacking volatile for *manual* type checking, at the cost of frustrating the optimizer? Andrei Alexandrescu once advocated that approach for *automatic* type checking in a famous article (http://drdobbs.com/cpp/184403766). I believe the shortcomings have been thoroughly discussed in comp.lang.c++. If you want to remind yourself, you could group the variable(s) and the mutex / semaphore in a structure, or name them similarly etc. > You are wrong in saying that 'volatile' has no place in multi-threading. > It is the correct way to go if you want to ensure that a value is e.g. > read/written just once even if it is used many times: It has no place in properly synchronized threaded programs. And it cannot guarantee the correctness of un-synchronized threaded programs (unless you assume non-SMP, non-hyper-threaded, Intely-type hardware -- *maybe*) > extern volatile int xval; // Written by other thread(s) > > void f (void) > { > int x; > > x = xval; > > // use x many times, it won't change. > } > > Without the 'volatile', the compiler is free to read > the memory value xval as many times as it wants, even > if it has a local copy, and it probably will do so if > you have many local variables. What does that accomplish? You're merely frustrating the compiler's ability to optimize. You're not achieving complete thread safety by *adding* volatile -- not on arbitrary hardware. If your code is completely thread-safe with volatile, it is also completely thread-safe (and faster) without volatile. Volatile does not offer any guarantees that cannot be later undone by the pipeline or CPU cache. -- Dan _______________________________________________ Linux-audio-dev mailing list [email protected] http://lists.linuxaudio.org/listinfo/linux-audio-dev
