Re: [AD] user-defined events

[Peter Wang <novalazy@xxxxxxxxxx>]

On 2008-09-27, Thomas Fjellstrom <tfjellstrom@xxxxxxxxxx> wrote:
> On Saturday 27 September 2008, Peter Wang wrote:
> > On 2008-09-27, Elias Pschernig <elias.pschernig@xxxxxxxxxx> wrote:
> > > No matter how many fields are added though, it never will be nice. How
> > > well would using just a single void *data work instead?
> > >
> > > It would be allocated whenever an event is emitted - but would it be
> > > clear to users where to free it? For example, what if the event is never
> > > received, or what if to event queues are receiving the events?
> > >
> > > I think we should make sure the above works, if it does not right now.
> > > It's the only way I can see this useful - if necessary, we might have to
> > > provide a callback pointer with a "cleanup_user_event" method which is
> > > passed the event when it goes out of servrice or something, so user code
> > > can free the data poiner there. (But maybe that's already working..
> > > didn't look too close yet and the example apparently just casts an int
> > > to a pointer.)
> >
> > It doesn't have any support for freeing user data in user events.
> >
> > Not sure what to do about it.
> 
> Little more serious this time, since your new event code will copy events 
> everywhere, shouldn't the user free the event data once they are done with it? 
> Just have to make sure the event queue still doesn't have a handle on it 
> before that. 

The problem is that it's not a deep copy.  Two events can point to the
same user data, e.g. if you register one event source with two queues.
Then the user doesn't know (easily) when the user data can be freed.

And as Elias pointed out:

1. You might try to emit an event, but the event source isn't connected
to any queues, or all the queues are full.  Then the user data needs to
be freed immediately.

2. Event queues might be non-empty when they are destroyed so you might
need to free some user data then too.

Having a cleanup_user_event callback would imply reference counting
again.  Can we just say that user events should not contain pointers
to dynamically allocated structures, and if they choose to ignore that
they need to take into account the points above?

Peter