accessing the shell from PD

Nice one, it probably works just fine, the issue is that time() will call into the Linux kernel and as it is called from within the audio thread it breaks the real-time guarantees of Bela, hence the "mode switch" warning.

You have several options:

• start up a thread and call time() in there, have the audio thread print the result that was obtained from the other thread. If you don't want the other thread to be polling all the time, you'll need to trigger a time() call in the other thread when you receive a bang, then set a pd timer (I believe that's what it's called?) a few ms later that reads the up-to-date result of time when it's ready and sends it out of the output. This is the most generally-applicable version, which will also improve real-time friendliness of your external when running outside Bela. It may get complicated because on a multi-core processor you'll need some way of guaranteeing memory coherence so that the audio thread never sees a corrupted version of the time and has a way of knowing for sure whether it's been updated.
• instead of calling time(), call __wrap_time(). This is a function provided by Xenomai's libcobalt that is safe to call from the audio thread. Before you call it you may have to declare it (depending on your compiler's options):

  extern time_t __wrap_time(time_t *tloc);

  I don't think you have to change your linker flags. You should be able to build it like this and at runtime it will be resolved because libcobalt is already loaded by the Bela binary.
• leave the external exactly as it is. Bang it only once at the beginning of the script to get the time, ignore the mode switch message. Then use bangs to the [timer] object to measure elapsed time while the program is running. I appreciate this may have accuracy implications for long running patches (which you can obviate to by resetting the timer every so often) and if you need the year/month/day/hour/min/sec the previous option may be easier.

The factoring out of x->getTime() looks good.

There are some errors in the file, it won't build as is. Among others, you are forward-declaring __wrap_time() but then using __wrap_clock_gettime().

rph-r I confess I'm not sure I understood everything I did with xenomai, but I'll dig that...

__wrap_SOMEFUNCTION() calls the Xenomai version of SOMEFUNCTION (if available). The forward declaration (extern someret __wrap_SOMEFUNCTION(someargs)) is a lazy workaround so that you don't need to find the correct file to include and add it to the include path or modify the #include statements).

Now the issue we encountered is that Xenomai's time() or clock_gettime(CLOCK_REALTIME, ...) gives unexpected results (it probably reports the time since boot?). This is probably a bug in Xenomai (or a feature, but I can't find it docuemented) and propagates to the use of time() - which also uses CLOCK_REALTIME under the hood. By trial and error I figured out that CLOCK_HOST_REALTIME yields correct results. However, time() doesn't support specifying the clock, so we have to use clock_gettime(CLOCK_HOST_REALTIME, ...) instead, and extract the seconds from the value it returns. As CLOCK_HOST_REALTIME is defined in the Xenomai include, we once more work around it by hardcoding its value 32.

Now that things work, if you wanted to tidy things up in a way that your external can be built on any platform but picks up Xenomai includes and functions if available, you could do the following:

• add this line to the Makefile before including the pdlibbuilder makefile:

  CPPFLAGS=-I/usr/xenomai/include/cobalt -I/usr/xenomai/include

  This will tell it to look for the Xenomai includes and it will pick up time.h from there, declaring __wrap_clock_gettime() and defining CLOCK_HOST_REALTIME
• then in the C file you can do simply:

  #ifdef CLOCK_HOST_REALTIME
    // Xenomai clock, if available
    __wrap_clock_gettime(CLOCK_HOST_REALTIME , &tp);
  #else
    clock_gettime(CLOCK_REALTIME , &tp);
  #endif

This is a bit nicer and doesn't require much platform detection: on most systems, /usr/xenomai/ won't exist and so the include paths you set in CPPFLAGS will be ignored. Note: in principle a change like this would also require to add a library to the list of shared libraries to link the external against. However, we rely on the host (the Bela core in this case) to load that library for us. That's not as neat, but it allows to keep things simple in the Makefile.

Further refactoring could see using a class_addanything instead of individual methods for the various year, month, day etc, but as that's so poorly documented (i.e.: undocumented) in Pd I don't blame you.

It would probably look something like:

void ahora_anything(t_ahora *x, t_symbol* s, int argc, t_atom* argv)
{
  x->get_time(x);
  if(s == gensym("year"))
    outlet_float(x->year_out , 1900+x->instant.tm_year) ;
 else if(s == gensym("month"))
    outlet_float(x->month_out , 1+x->instant.tm_mon) ;
else if ...
//
else
  pd_error(x, "no method for %s\n", s->s_name);
}


// in ahora_setup(), instead of all the individual ahora_addmethod() calls:
class_addanything(ahora_class, (t_anymethod)ahora_anything);

Not sure if's worth anyone's time, but it would look nicer -- if it works. I hope the bang method still works, otherwise you need to add it as a case in the chain of ifs in ahora_anything().

Btw, you may also want to make the creation argument a symbol instead of a float and convert that to a number internally before storing it into x->whichout, that would make it more user-friendly.