Re: [Jack-Devel] JACK midi events lost on 64 bit machines

Hi Carlos,

On Sun, Jun 26, 2011 at 5:50 PM, Carlos Eduardo da Fonseca
<[hidden]> wrote:

> I have an application using JACK MIDI under Linux. It works perfectly on 32
> bit machines, but on my new one, which is 64 bits, I get event loss all the
> time.

Your message doesn't mention which version of JACK you are running on
your new machine, or which version of JACK you ran on your old
machine.  That data might be helpful.

> #define MidiMessageSize 3
>
> jack_client_t* client;
> jack_port_t* midi_out_port;
>
> char current_message[MidiMessageSize];
>
> jack_ringbuffer_t* midi_rb;
>
> int process (jack_nframes_t nframes, void *arg) {
>   void* output_buffer = jack_port_get_buffer(midi_out_port, nframes);
>   jack_midi_clear_buffer(output_buffer);
>   int read_space = jack_ringbuffer_read_space(midi_rb);
>   if (read_space == 0) return 0;
>   int i;
>   for (i = 0; i < read_space; i += MidiMessageSize) {
>     jack_ringbuffer_read(midi_rb, current_message, MidiMessageSize);
>     jack_midi_event_write(output_buffer, i/MidiMessageSize,
> 			   (jack_midi_data_t*)current_message, MidiMessageSize);
>     if (jack_midi_get_lost_event_count(midi_out_port)) {
>       printf("Lost\n"); /* Never get here on 32 bit but always on 64 */
>     }
>   }
>   jack_ringbuffer_reset(midi_rb);
>   return 0;
> }

If your period size is smaller than the amount of messages retrieved
from your ringbuffer during a process call, then your timestamp (i /
MidiMessageSize) will become out-of-bounds.  However, at least with
JACK 2, the error wouldn't increment the lost event count.

With JACK 2, events may be considered "lost" 1.) during mixdown, when
two or more MIDI buffers are having their contents transferred to
another MIDI buffer and a MIDI message can't be transferred to the
MIDI buffer, and 2.) when space cannot be reserved in the MIDI buffer
for your event when calling 'jack_midi_event_write' or
'jack_midi_event_reserve'.  You can find the relevant code here:

    http://trac.jackaudio.org/browser/jack2/trunk/jackmp/common/JackMidiAPI.cpp

... and here:

    http://trac.jackaudio.org/browser/jack2/trunk/jackmp/common/JackMidiPort.cpp

The second case is the only case that matters in this situation, as
you're calling 'jack_midi_clear_buffer' on the output buffer at the
beginning of every process call, which will reset the lost event count
to zero. You can check for the second case by checking if the return
value of 'jack_midi_event_write' is -ENOBUFS.

FYI, 'jack_ringbuffer_reset' is not thread-safe.  Calling it in a
process callback is probably a bad idea.

-- 
Devin Anderson
devin (at) charityfinders (dot) com

CharityFinders - http://www.charityfinders.com/
synthclone - http://synthclone.googlecode.com/

1309140654.18988_0.ltw:2,a <BANLkTik145H8_vfKaTOKyYyB1ngOeW7CHg at mail dot gmail dot com>