Re: nss_ldap broken

On Fri, 26 Mar 2004, Sean McNeil wrote:

> 
> On Fri, 2004-03-26 at 14:51, Daniel Eischen wrote:
> > 
> > I think I made a comment about how you should always
> > prefix _pthread_foo() calls with 'if (__isthreaded)'.
> > 
> > When the thread libraries are initialized, then overrwrite
> > the function pointers in libc's thread jumptable.  If you
> > unload the library, libc still retains those pointers.
> 
> Yes, checking __isthreaded solves my problem.  nsdispatch.c has several
> calls to _thread* routines that should be protected by the __isthreaded
> test.
> 
> It would appear that even though RTLD_LOCAL is used on the dlopen, since
> nss_ldap.so.1 is linked with libc as well it causes those thread
> jumptable entries to be modified.  When dlclose is called the reference
> count to the thread library goes to zero, the library is unloaded, and
> the pointers are now no good.
> 
> This poses an interesting problem, though.  Since there is a dlopen in
> this it is possible for part of a function to have __isthreaded not set
> and part of it set.  For these routines the value of __isthreaded would
> need to be stashed.  Yet for the nss_atexit, since it is the opposite
> direction (dlclose), the variable would have to be tested directly.

Once something is threaded, I think it always has to be
threaded.  We don't support going back to unthreaded mode.

I'm not sure how all this works with nss_ldap (and perhaps
other modules) and why they are bringing in the threads
library.  What dependency is bringing in the threads library
in the first place?

-- 
Dan Eischen