Out-of-order Dll Teardown

[Ben Creech]

This question isn't exactly Kerberos-specific, but it is still pertinent
to this list.  In summary, I'm getting odd behavior from Windows on
teardown of KfW's krb5_32.dll.

As noted in athena\auth\krbcc\src\dll\cacheapi.cxx, Windows doesn't
follow the Dll dependency tree when tearing down Dlls with calls to
DllMain with DLL_PROCESS_DETACH.  For example, if A.exe loads B.dll
which in turn loads C.dll, Windows apparently sometimes tears down C.dll
before B.dll (A->C->B instead of A->B->C).

I say "apparently" because I have observed this in my own program. 
A.exe loads B.dll which delay-loads krb5_32.dll.  B.dll calls
krb5_init_context.  Things Happen.  A.exe exits.  Windows tears down
krb5_32.dll, which shuts down the ccache connection and frees its heap
(via the run-time library).  At this point, B.dll's DllMain is finally
called.  As part of the cleanup routine, B.dll calls krb5_free_context. 
Whoops, that's no longer a valid pointer!  Program crashes.

Of course, I can simply comment out my krb5_free_context since the
process is exiting anyway [1].  Alternately, I can make A.exe call a
cleanup function before it exits [2].

Incidentally, I didn't get this behavior until I started using debug
dlls from 2.2-beta-2; the downloadable 2.1 release build didn't do it. 
After poking through the code, I'm fairly satisfied that 2.2 added
nothing that would cause my problem, so it's probably the fact that I'm
using debug binaries now.

Can anyone tell me if my logic is correct?  I'm worried that I'm either
doing something wrong on my side, or there is a subtle bug in 2.2 beta 2
(is anyone testing it with delay-loading)?

Thanks,
Ben Creech
NCSU ITECS