[krbdev.mit.edu #7135] gssapi mechanism glue dlcloses objects potentially after they are already unloaded

[Sam Hartman via RT]

moonshot has consistently gotten a linker error on Linux at process exit
time. I traced it with a debugger today and figured out what seems to be
going on.

exit calls the library finalization functions for loaded libraries.
That calls gssint_mechglue_fini.  If PROGRAM_EXITING() returns false
(which it always does) that calls freeMechList() which indirectly calls
dlclose on loaded objects.

Unfortunately, moonshot depends circularly on libgssapi_krb5. The linker
chooses to break this dependency by finalizing moonshot first.
Then when we call dlclose, we get:

Inconsistency detected by ld.so: dl-close.c: 743: _dl_close: Assertion
`map->l_i nit_called' failed!


This is arguably a linker bug.

A few observations:

1) It would be nice to have a better definition of PROGRAM_EXITING for
glibc platforms.

2) There are reasonably good reasons for GSS mechanisms to call into
libgssapi_krb5 even though they are loaded by it. 

Finally, I suspect that gssint_mechglue_fini may tend to be useless if a
mechanism that calls into the mechglue is loaded.  Consider what
happens.  The cases where fini matters is where the gss library would
like to be unloaded.  So, an application dlopens a plugin that loads
gssapi_krb5.  That eventually opens moonshot, which creates a second
reference to libgssapi_krb5.

The application dlcloses the plugin, removing one reference to
libgssapi_krb5.  The linker really needs to garbage collect or otherwise
walk the graph in order to determine that it can get rid of the
moonshot-gssapi subgraph.  So on Linux, it looks a lot like this code
crashes at process exit and is likely to be useless otherwise.