dlopen and C++ libraries (vtables)

[Matthew Chapman]

Hi,

I've run into an issue with dlopen and loading C++ libraries.  I
apologise if this is a FAQ and "can't be done" or "is easily solved".
This is happening deep inside a third-party commercial tool, so it
would be nice if I could feed back some advice for dealing with it.


Consider a program (main) that depends on a plugin (libmine) that then
happens to depend on some system library written in C++ (e.g.
libstdc++ or libboost or libxerces-c).

If main is linked at build time with libmine.so, it seems this always
works as expected.  Initializers are called in the order
/usr/lib64/libxerces-c.so, ./libmine.so, ./main; and finalizers in the
reverse order ./main, ./libmine.so, /usr/lib64/libxerces-c.so.


However, if main loads libmine.so with dlopen(), things can go bad.
During relocation processing, the dynamic linker decides:
      3199:     file=./libmine.so [0];  needed by
/usr/lib64/libxerces-c.so [0] (relocation dependency)
(_ZTVN11xercesc_2_713XSerializableE)

i.e. there is a false circular dependency created where libxerces-c.so
requires ./libmine.so.  The problem is that the same vtables and
typeinfo structures will often be weakly defined in both libraries,
e.g.:

./libmine.so:
0000000000200df0 V _ZTVN11xercesc_2_813XSerializableE
/usr/lib64/libxerces-c.so:
00000000005797a0 V _ZTVN11xercesc_2_813XSerializableE

and it seems that the first one is always chosen (load order rather
than dependency order?).

The problem then is that - due to the false dependency - the libraries
then sometimes get unloaded in the wrong order, causing symbol
resolution errors (or worse) in destructors.  I can come up with a
test case if needed.

As previously noted this doesn't happen when the library is loaded as
part of the normal dependencies, which must use a different resolution
order.


Are there any ways of getting around this behaviour?  I've tried
various dlopen flags with no success so far.  I would argue that it if
it's possible to make libmine.so load correctly at startup, it should
also be possible to make it work with dlopen()...

Thanks,
Matt