Hi,
When writing an application that records midi (and e g saves it to disk), ultra-low latency isn't really needed, all we need to know is exactly when the midi message came in. The application can then wake up once a second or so, to write the incoming data, including an accurate timestamp for every event, to disk.
As far as I can see, the rawmidi interface does not support such a feature at all. There is a snd_rawmidi_status syscall, but its timestamp field is not even filled by the kernel (!). But even if that was fixed, it would not fix the problem as there could be several midi events in the buffer with different timestamps.
You could use the seq interface, it does support timestamps, but I can see at least two potential problems with this:
First, the seq code runs in a work queue, not in the actual IRQ. This means that midi event is timestamped too late, especially so if the work is delayed for some reason.
Second, seq hard-codes the timestamp type to monotonic - there is no monotonic_raw, so the timestamp would be affected by NTP changes.
Also, the timespec uses 32-bit for sec and nsec, but I suspect this is less of a problem (unless people constantly record midi for sixty years or so...).
It seems like it could be useful to have a rawmidi mode which automatically prefixes every incoming message with a timestamp. This would replace the current "raw" midi buffer with some type of header "length, timestamp, data" or so. Arguably this would make the rawmidi slightly less raw, but anyway. Does this make sense? Should I attempt to write something? Or would it be better to try to use (and potentially improve) the seq code?
// David