I think I can answer my own question ...

[Bryan Ischo <bryan at ischo dot com>]

I managed to dig into the glibc source code and found an interesting 
function:

int
__register_atfork (prepare, parent, child, dso_handle)
     void (*prepare) (void);
     void (*parent) (void);
     void (*child) (void);
     void *dso_handle;

I used my LD_PRELOAD trick to define this to a dummy function that does 
nothing other than print out a message saying that the call is being 
ignored.

When I run my program using this dummy __register_atfork, it seems to 
fix the problem.  It appears that __register_atfork is called twice.  I 
have managed to get a stack trace from both of these; they are:

(gdb) where
#0  0x00007fa82c8b8b40 in __nanosleep_nocancel () from /lib/libc.so.6
#1  0x00007fa82c8b8a01 in sleep () from /lib/libc.so.6
#2  0x00007fa82e66681c in __register_atfork () from /home/bji/nvidia_fix.so
#3  0x00007fa82d66752a in ?? () from /usr/lib/libGL.so.1
#4  0x00007fa82d6414ec in ?? () from /usr/lib/libGL.so.1
#5  0x00007fa82e875902 in call_init () from /lib/ld-linux-x86-64.so.2
#6  0x00007fa82e875a2a in _dl_init_internal () from 
/lib/ld-linux-x86-64.so.2
#7  0x00007fa82e8685fa in _dl_start_user () from /lib/ld-linux-x86-64.so.2
#8  0x0000000000000002 in ?? ()
#9  0x00007fff5b43db1c in ?? ()
#10 0x00007fff5b43db47 in ?? ()
#11 0x0000000000000000 in ?? ()
(gdb) c
  C-c C-c
Program received signal SIGINT, Interrupt.
0x00007fa82c8b8b40 in __nanosleep_nocancel () from /lib/libc.so.6
(gdb) where
#0  0x00007fa82c8b8b40 in __nanosleep_nocancel () from /lib/libc.so.6
#1  0x00007fa82c8b8a01 in sleep () from /lib/libc.so.6
#2  0x00007fa82e66681c in __register_atfork () from /home/bji/nvidia_fix.so
#3  0x00007fa82d6414b1 in ?? () from /usr/lib/libGL.so.1
#4  0x00007fa82e875902 in call_init () from /lib/ld-linux-x86-64.so.2
#5  0x00007fa82e875a2a in _dl_init_internal () from 
/lib/ld-linux-x86-64.so.2
#6  0x00007fa82e8685fa in _dl_start_user () from /lib/ld-linux-x86-64.so.2
#7  0x0000000000000002 in ?? ()
#8  0x00007fff5b43db1c in ?? ()
#9  0x00007fff5b43db47 in ?? ()
#10 0x0000000000000000 in ?? ()

(I got the stack trace by putting a sleep(10) inside my dummy 
__register_atfork and then breaking into the debugger; a cheesy but 
effective technique).

Can someone explain __register_atfork to me?  Why would NVidia call that 
function directly instead of pthread_atfork()?  And is this really valid 
for NVidia to be doing?

Is there any better way to disable these (in my opinion, broken) calls 
into __register_atfork than using an LD_PRELOAD trick?

Thank you!
Bryan