NRT thread can block disk I/O thread

[Spacechild1]

The disk I/O thread in DiskIO_UGens.cpp locks the NRT thread when it wants to write data from the buffer to the disk. AFAICT, this is only necessary because the user might accidentally free the buffer concurrently (which shouldn't really happen).

I understand that (ab)using Server sound buffers for disk I/O is convenient because it already handles the opening and closing of soundfiles. However, it also means the disk I/O thread is implicitly tied to the NRT thread, which kind of defeats the whole purpose of having a dedicated disk I/O thread in the first place...

Note that asynchronous commands executed on the NRT thread can take an arbitrary amount of time!

This causes problems such as this: https://git.iem.at/pd/vstplugin/-/issues/58

[joshpar]

This lock was added, I think, to make sure the disk io ugens worked in NRT mode. /* Josh Parmenter www.realizedsound.net/josh */

…

On Jul 17, 2021, at 16:12, Christof Ressi ***@***.***> wrote:  The disk I/O thread in DiskIO_UGens.cpp locks the NRT thread when it wants to write data from the buffer to the disk. AFAICT, this is only necessary because the user might accidentally free the buffer concurrently (which shouldn't really happen). I understand that (ab)using Server sound buffers for disk I/O is convenient because it already handles the opening and closing of soundfiles. However, it also means the disk I/O thread is implicitly tied to the NRT thread, which kind of defeats the whole purpose of having a dedicated disk I/O thread in the first place... Note that asynchronous commands executed on the NRT thread can take an arbitrary amount of time! This causes problems such as this: https://git.iem.at/pd/vstplugin/-/issues/58 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[Spacechild1]

No, DiskIn and VDiskIn contain different code paths for RT and NRT synthesis. The latter doesn't use the disk I/O thread at all, instead it reads from the file synchronously.

---

BTW, I find it strange that DiskOut does use the disk I/O thread in NRT mode. There is basically no synchronization between the main thread and the I/O thread, so if the latter can't catch up with the former, it would drop audio... I think it should write to disk synchronously instead.

[joshpar]

I'm surprised. I've used DiskOut a LOT for large NRT files without any problems. Regardless - if you propose any changes, please make sure reading and writing for those UGens in NRT still works. Thanks!

[Spacechild1]

I've used DiskOut a LOT for large NRT files without any problems

That's interesting to know! In practice it's not a problem if the buffer is sufficiently big, but it's definitely not safe.

please make sure reading and writing for those UGens in NRT still works

Sure, good point!