[Halld-offline] collisions among static and dynmaic (plugins) libraries

[David Lawrence]

Hi Richard,

  I spoke with Paul since I know we ran into this some time ago and found a solution. Digging through
the github logs, here is where Paul implemented the solution we came up with:

commit 61500561a96d5cb8d471c18bdb38d911611f3da9
Author: Paul Mattione <pmatt at jlab.org>
Date:   Tue Mar 17 15:35:54 2015 +0000

    Make global static variables in DTranslationTable have function scope instead. Return these by reference from these container functions, and access only by these functions.
    This fixes the crash-at-end-of-program problem (double-free, corrupted-memory) mentioned in the rev 17681 hd_root commit. Remove abort() call from hd_root.

For example, at the top of DTranslationTable.cc there is this:

// Use one translation table for all threads
pthread_mutex_t& DTranslationTable::Get_TT_Mutex(void) const
{
	static pthread_mutex_t tt_mutex = PTHREAD_MUTEX_INITIALIZER;
	return tt_mutex;
}


Previously, the “tt_mutex” variable was global so when a plugin that had been linked with
the TTAB library was attached (dlopen), it automatically linked this. When the plugin is 
detached (dlclose) it was treated it like it went out of scope, deleting it. Then, the same thing 
happened at program exit leading to a double-free leading to a crash.

By replacing the global with a static local variable inside a globally accessible method, one
avoids this. So, you will need to hunt down where in the code a global variable is being used
and wrap it inside a global getter method as is done above in order to avoid these crashes.

Regards,
-David



> On Jul 13, 2016, at 3:11 PM, Paul Mattione <pmatt at jlab.org> wrote:
> 
> It’s only #include’d in danarest.  So, this problem would apply to every single class that’s #include’d into a plugin (such as DChargedTrack), so it’s unavoidable.  
> 
> - Paul
> 
> On Jul 13, 2016, at 3:05 PM, Richard Jones <richard.t.jones at uconn.edu> wrote:
> 
>> Hello Dave,
>> 
>> I have run into a nasty problem with segfaults at program exit. This seems to be related to the appearance of duplicate instances of some classes (eg. DEventWriterREST) that are defined both in the plugin (danarest.so) and the static library that is linked into hd_root (libHDDM.a). What is your policy with regard to classes that are used both in the static library and in the base code? 
>> 
>> BTW there is the related question of why the DEventWriterREST is duplicated in two places in the tree, once in the Utilitiles/plugins/danarest and once in the libraries/HDDM area.
>> 
>> -Richard Jones
>> _______________________________________________
>> Halld-offline mailing list
>> Halld-offline at jlab.org
>> https://mailman.jlab.org/mailman/listinfo/halld-offline
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.jlab.org/pipermail/halld-offline/attachments/20160713/ee3ee8fd/attachment-0002.html>