agent, dirmngr, scd: npth_init must be after fork.

* agent/gpg-agent.c (thread_init_once, initialize_modules): New.
(main): Make sure no daemonizing-fork call after npth_init, and no npth
calls before npth_init, with care of npth calls by assuan hooks.
* dirmngr/dirmngr.c (thread_init): New.
(main): Make sure npth_init must not be called before daemonizing fork.
* scd/scdaemon.c (main): Likewise.

--

It is simply the best for nPth not to allow the daemonizing fork after
npth_init, because semantics and implementations of forked child process
in a threaded application is a difficult corner case.

GnuPG-bug-id: 1779
Signed-off-by: NIIBE Yutaka <gniibe@fsij.org>

scd/scdaemon.c

@@ -422,8 +422,6 @@ main (int argc, char **argv )
   i18n_init ();
   init_common_subsystems (&argc, &argv);
 
-  npth_init ();
-
   ksba_set_malloc_hooks (gcry_malloc, gcry_realloc, gcry_free);
 
   malloc_hooks.malloc = gcry_malloc;
@@ -724,6 +722,8 @@ main (int argc, char **argv )
       }
 #endif
 
+      npth_init ();
+
       /* If --debug-allow-core-dump has been given we also need to
          switch the working directory to a place where we can actually
          write. */
@@ -861,6 +861,8 @@ main (int argc, char **argv )
 
       /* This is the child. */
 
+      npth_init ();
+
       /* Detach from tty and put process into a new session. */
       if (!nodetach )
         {