Monday, April 13, 2009

Shared Libraries

Shared libraries, dynamic linked libraries, .so, .dll, .dylib, whatever you might call them seem to be a bit of a black art of programming. Especially when you mix C++ into the picture.

One key part of PAL is the self-registering plugin system. This allows extra functionality to be extended to the system on demand. It was very useful to allow the development of various simulation systems where fixes or alterations to the physics system can be loaded on demand.

It also greatly simplifies the whole process of creating/managing objects. The big catch is getting it to work in a portable cross platform manner.

The basics of the system is this:
Abstract factory which will have plug-ins self register via either a static constructor, or add themselves to a list of available constructors that can be retrieved via a shared object function call.

For staticly constructed objects, its not a big deal, just have a single static version of each object which registers itself with the factory. This works fine as long as those static variables are initialized. Unfortunately, the Microsoft compiler does not link those symbols from a static library, unless a function in the static library is explicitly called. So your options are to link all the objects to your program, or have a 'stub' in the static library, that you are forced to call to populate all the objects. PAL follows this latter approach. (Note, if you put that static stub into a separate header file, you can save yourself a lot of include header problems)

For dynamic libraries, well, it's not such a big deal for Windows. Under MSVC you just create a DLL project, and load the library with LoadLibrary and GetProcAddress. Ofcourse you need to remember to '__declspec(dllexport)' the functions you need to access, or write the def file (far too complicated). Then make sure the compiler exports the symbols.

MacOSX didn't pose great problems either, again, just use dlopen and dlsym. Remember to compile with '-dynamiclib'.

That brings my to GCC/Linux. This one was a bit tricky, and I could only solve this with the kind help of B. Watson over on #slackware. The problem with GCC is the way it handles run time type information. You need RTTI for dynamic_cast's etc, to work. This is relatively critical for design patterns such as abstract factories and virtual inheritance, which are heavily employed in PAL.

Under Linux to open a shared object, dlopen and dlsym will do, however you need to specify RTLD_GLOBAL, otherwise the type info will be stored locally and not globally. Now this normally doesn't cause a problem, but with GCC class types are compared by pointer references and not strings, so of course, with each object having its own local address space, the code just wont work!

That fixes one problem, the other is getting the symbols to export. By default GCC won't export all the symbols that you need. To force GCC to do this we need to pass through to the linker from the compiler the '-rdynamic' flag. But, alas, we are hampered again, we need to actually make sure the symbols are visible. GCC has introduce the ability to have private/local symbols, and if only one person specifies something to be private then the whole subsystem becomes private. No major dramas though, just specify '__attribute__ ((visibility("default")))' everywhere.

So now finally your ready to compile a linux shared object with position independent code:
-fPIC -shared -rdynamic
But it still doesn't work!

Yes, for some reason I don't know, you need to make sure your MAIN program is also compiled with the '-rdynamic' flag.

I guess the lesson for the end of the day was, '-rdynamic' is your friend.

Heres a useful snippet that summerizes all this:

REMEMBER: Export ALL symbols (shared or not!)
Linux: RTLD_GLOBAL & rdynamic!



#if defined (OS_WINDOWS) || defined(WIN32)
# define DYNLIB_HANDLE hInstance
# define DYNLIB_LOAD( a ) LoadLibrary( a )
# define DYNLIB_GETSYM( a, b ) GetProcAddress( a, b )
# define DYNLIB_UNLOAD( a ) !FreeLibrary( a )

struct HINSTANCE__;
typedef struct HINSTANCE__* hInstance;

#elif defined (OS_LINUX) || defined(OS_OSX)
# define DYNLIB_HANDLE void*
# define DYNLIB_LOAD( a ) dlopen( a, RTLD_LAZY|RTLD_GLOBAL )
# define DYNLIB_GETSYM( a, b ) dlsym( a, b )
# define DYNLIB_UNLOAD( a ) dlclose( a )
#endif


To compile:

OSX:
main:
g++ main.cpp -o main
DLL:
g++ shared.cpp -o shared.dylib -dynamiclib

Linux:
main:
g++ main.cpp -o main -ldl -rdynamic -fPIC
g++ shared.cpp -o shared_single.so -fPIC -shared -rdynamic


And would you believe it took me almost 3 hours to figure this all out. (At least I didn't end up having to write my own RTTI system like the last guy.

You can check out the PAL self-registering abstract pluggable factory in the code. (see also the test programs and makefiles).

No comments: