Hi Mat,

 >>> be optimized (use a bitmap to flag which ones to free), but there's
>>> another problem: watch_list is indexed by an fd which was likely closed
>>> by the data->destroy() callback. It's possible for the fd to get re-used
>>> and a new watch to be created on that fd even before the data->destroy
>>> callback returns.
>>
>> But then won't you be calling the watch's callback erroneously anyway?
>> If this is a concern, then all new watches need to be postponed until
>> after the epoll processing has been completed.
>>
>
> No, the stale data is a pointer to watch_data not an fd value. By
> deferring the free we can make sure the event loop is calling something
> valid. The contents of watch_data can be reset however we need to. The
> problem is with pointer values stored in watch_list getting clobbered
> before they get used to free memory.

Now I get it.  I'm good with the proposed solution, but can we avoid 
having dummy callbacks?  Why not simply set the callback to zero (or set 
some flag) and not call it.

Regards,
-Denis