Long story short:

After following a child, detaching from the parent,

('set follow-fork-mode child' + 'set detach-on-fork on')

here in this bit,

 infrun.c:resume():
  {
    ....
    follow_fork ();
    ...
    tp->stop_signal = TARGET_SIGNAL_0;
  }

... `tp' is no longer in the thread list (it was pointing at a thread
of the parent process, which we've detached from, hence no longer
in the thread list), so if the assignment above doesn't crash, it ends
up writing to who-knows-where.

With some local changes I was making, sometimes, `tp' happened to be left pointing
at linux_nat.c:lwp_list, and so that assignment above ended up clearing
lp->waitstatus.kind (of the first lwp in the list), which resulted in
GDB considering that the child process had
exited (because TARGET_SIGNAL_0 == TARGET_WAITKIND_EXITED).

This should fix intermittent foll-fork.exp foll-vfork.exp
fork-child-threads.exp failures.

Checked in.

-- 
Pedro Alves