We hit an assertion in ld.so about every 6000 runs of the cluster manager
slurmstepd on SiCortex hardware.  This is evidently the same bug as

http://www.redhat.com/archives/phil-list/2003-December/msg00008.html

Evidently it has reappeared because of ld.so consistency checking and 
because our chip (6 way SMP at 500 MHz) has a wider window of vulnerability.

We've adapted the previously reported test case so that it fails about half the
time (attached) and developed a patch (attached) that resolves the problem.

The test case doesn't fail for us on opterons, the only other systems we have
available.

The failure message we get on our machines is:

Inconsistency detected by ld.so: dl-open.c: 215: dl_open_worker: Assertion
`_dl_debug_initialize (0, args->nsid)->r_state == RT_CONSISTENT' failed!

Details:

If a thread happens to hold dl_load_lock and have r_state set to RT_ADD or
RT_DELETE at the time another thread calls fork(), then the child exit code
from fork (in nptl/sysdeps/unix/sysv/linux/fork.c in our case) re-initializes
dl_load_lock but does not restore r_state to RT_CONSISTENT. If the child
subsequently requires ld.so functionality before calling exec(), then the
assertion will fire.

The patch acquires dl_load_lock on entry to fork() and releases it on exit
from the parent path.  The child path is initialized as currently done.
This is essentially pthreads_atfork, but forced to be first because the
acquisition of dl_load_lock must happen before malloc_atfork is active
to avoid a deadlock.