I think you can get away with sharing a lock between two high-prioroty processes as long as neither one holds the lock for more than a small amount of time and if the OS can be counted on to give control to a real-time process quickly once it becomes runnable (i.e., if it's blocked on a lock, once that lock is released).
The situation I don't know about is this: if Pd's main thread failed to get the lock, so that control (presumably) passed back to the other thread that had the lock, how much time can pass before the other thread blocks on something so that control (again presumably) gets passed back to the main thread?
But anyway, since neither thread holds onto the lock for more than a few lines of C code (with no system calls) it's probably blue-moon rare that the scheduler interrupts one thread right in the middle of a critical section and passes control to the other one that then blocks. So this is essentially untested.
Threads can never be used confidently in a real-time situation. But I don't see any reasonable way without them to implement readsf~/writesf~, so there we are...
cheers Miller
P.S. one can issue non-blocking reads/writes, but there's also "open" which is much more likely to hiccup than "read", and I don't know of any async open call in any OS.
On Tue, Oct 06, 2015 at 03:10:31AM +0000, Jonathan Wilkes via Pd-list wrote:
- One thing I noticed is that the article you cited seems to focus
on tasks not critical to the computation/delivery of audio samples.For example, if your program were blocking or locking in order to do a GUIupdate. But here, the data must arrive in time to compute the next block. If ittakes too long to read the next portion of the sound file, then you're going to geta glitch. But I'm not sure I really grasp how locking works, nor really the whole file i/oprocess in general.
Here's a naive question: why can't you just tell the OS to treat the file asif it were a non-blocking socket, add the fd to Pd's event loop with sys_addpollfn, and then receive the incoming data to the relevant function?(Warning: some or all of the above may technically be gibberish...)
-Jonathan
On Monday, October 5, 2015 10:01 PM, Robert Esler <robert@urbanstew.org> wrote:I’m trying to understand why readsf~ and writesf~ work so well.
I’m particularly referencing Ross Bencina’s article: http://www.rossbencina.com/code/real-time-audio-programming-101-time-waits-f... and his subsequent paper, http://www.rossbencina.com/static/writings/File_IO_ACMC2014_Bencina.pdf
If you are not into asynchronous message passing and lock-free queueing then I’ll summarize the articles briefly:
When engaging in file I/O (e.g reading from or writing to an audio file) do not use locks or blocking. He goes on to say that this can lead to priority inversion, unbound execution time and “scheduler paranoia”.
This is all absolutely true in my experience in the audio jungle.
Pd’s async file I/O objects (readsf~ and writesf~) use both locks and blocking via a mutex and the pthread_cond_signal and pthread_cond_init functions. Look at the source code file d_soundfile.c for more details. The gist of it is that these objects have two threads. One parent thread that sends the data to the dsp scheduler, and a child thread that grabs the data from the file, and subsequently the child signals the parent when it has more data.
Based on Bencina’s paper, readsf~ and writesf~ could (should?) glitch and may not be real-time safe.
My questions are:
Have I completely misunderstood d_soundfile.c and it is actually entirely safe. If so, why is it safe?
Why doesn’t Pd glitch more often when using these objects?
Does Pd need lock-free message queueing for such inter-thread communication?
Has anyone ever “broken” these objects or experienced glitching?
Thanks for the extra brain power. -R
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list