WIP: Add a COB_LOAD_GLOBAL config option to modify dlopen behavior

[florianschmidt1994]

This PR adds a config option called COB_LOAD_GLOBAL which allows the flags of dlopen used in call.c to be changed between RTLD_GLOBAL and RTLD_LOCAL (see https://linux.die.net/man/3/dlopen)

Motivation

This flag is motivated by an ambition to be able to write a multi-threaded c program, that then loads libcob with dlmopen in one namespace per thread, effectively allowing us to call the same libcob / Cobol module from multiple threads in a "thread-safe" way. The image below shows the a very simplified intended call sequence

Because there is a "bug" in dlopen that stops us from calling dlopen("module", RTLD_GLOBAL), this PR introduces an option to be able to change this flag to RTLD_LOCAL at runtime.

As at glibc 2.24, specifying the RTLD_GLOBAL flag when calling dlmopen() generates an error. Furthermore, specifying RTLD_GLOBAL when calling dlopen() results in a program crash (SIGSEGV) if the call is made from any object loaded in a namespace other than the initial namespace.

Docs

# Environment name:  COB_LOAD_GLOBAL
#   Parameter name:  load_global
#          Purpose:  tell the system loader to provide symbols in CALLed
#                    programs globally, allowing symbols to be found for linked
#                    libraries later
#             Type:  boolean
#             Note:  COBOL CALLs will always find symbols in already CALLed or
#                    pre-modules; this setting is mostly an advise to the system,
#                    not all systems are capable of loading libraries global/local
#          Default:  false
#          Example:  load_global true

Implementation

• Change lt_dlopen from a macro definition to an actual function, which internally calls dlopen with the flags based on the configuration
• Add cob_load_global / COB_LOAD_GLOBAL as a configuration
• Add documentation to runtime.cfg about COB_LOAD_GLOBAL

[GitMensch]

Apart from the other PR's review to that branch, the change here makes the current scenario even more confusing.
I'd like to fix this in this PR (or up-front upstream).

After the changes:

• instead of calling lt_dlopen we call a new static function cob_dlopen
• this has the current code of this PR in for the case of dlopen usage
• for the WIN32 variant it uses what is currently used there
• for the libtool variant it uses lt_dlopenadvise() with an advise matching the runtime setting

@florianschmidt1994 can you please inspect the libtool variant (to test this you can simply undefine USE_LIBDL in config.h)?
I think I'll work on the cob_dlopen stuff late-evening.

[florianschmidt1994]

Yup, I'll check it out. Basically something like this, right?

static void* cob_dlopen(const char* filename) {
	#ifdef _WIN32
	if (x == NULL) {
		return GetModuleHandle (NULL);
	}
	return LoadLibrary(x);	
	#elif	defined(USE_LIBDL)
	int flags = cobsetptr->cob_load_global 
		? RTLD_LAZY | RTLD_GLOBAL
		: RTLD_LAZY | RTLD_LOCAL;

	return dlopen(filename, flags);
	# else
	// TODO: initialise lt_dladvise
	lt_dlopenadvise(filename, /*advise here*/);
}

[florianschmidt1994]

@GitMensch I've tried out your suggestion in the latest commit and temporarily defined lt_dlopen as cob_dlopen to try it out. I've also added some indentation to the macros, does that conform with your style or should I remove them again?

[GitMensch]

nice update, we may be able to push that soon

Note: we still need the testcase addition - does not need to be something that benefit from this as this is hard to make portable - just do a simple CALL and CANCEL. We should have an easy one for COB_LOAD_CASE

[GitMensch]

please make the advise static and clean it up in call.c's cleanup function (depending on its preprocessor macro, of course), you can also init it in the init function - just set the advise flag here (the setting for both dlopen and lt_dlopen will be moved out to a config function later - you're welcome to add it (see screenio.c, this may be even explained in the video :-) and call it from the config side if this var or COB_LIBRARY_PATH was adjusted

[GitMensch]

After initial testing - please combine those tests into one, possibly named "CALL/CANCEL and COB_LOAD_GLOBAL"