elf: Always call destructors in reverse constructor order (bug 30785)

The current implementation of dlclose (and process exit) re-sorts the
link maps before calling ELF destructors.  Destructor order is not the
reverse of the constructor order as a result: The second sort takes
relocation dependencies into account, and other differences can result
from ambiguous inputs, such as cycles.  (The force_first handling in
_dl_sort_maps is not effective for dlclose.)  After the changes in
this commit, there is still a required difference due to
dlopen/dlclose ordering by the application, but the previous
discrepancies went beyond that.

A new global (namespace-spanning) list of link maps,
_dl_init_called_list, is updated right before ELF constructors are
called from _dl_init.

In dl_close_worker, the maps variable, an on-stack variable length
array, is eliminated.  (VLAs are problematic, and dlclose should not
call malloc because it cannot readily deal with malloc failure.)
Marking still-used objects uses the namespace list directly, with
next and next_idx replacing the done_index variable.

After marking, _dl_init_called_list is used to call the destructors
of now-unused maps in reverse destructor order.  These destructors
can call dlopen.  Previously, new objects do not have l_map_used set.
This had to change: There is no copy of the link map list anymore,
so processing would cover newly opened (and unmarked) mappings,
unloading them.  Now, _dl_init (indirectly) sets l_map_used, too.
(dlclose is handled by the existing reentrancy guard.)

After _dl_init_called_list traversal, two more loops follow.  The
processing order changes to the original link map order in the
namespace.  Previously, dependency order was used.  The difference
should not matter because relocation dependencies could already
reorder link maps in the old code.

The changes to _dl_fini remove the sorting step and replace it with
a traversal of _dl_init_called_list.  The l_direct_opencount
decrement outside the loader lock is removed because it appears
incorrect: the counter manipulation could race with other dynamic
loader operations.

tst-audit23 needs adjustments to the changes in LA_ACT_DELETE
notifications.  The new approach for checking la_activity should
make it clearer that la_activty calls come in pairs around namespace
updates.

The dependency sorting test cases need updates because the destructor
order is always the opposite order of constructor order, even with
relocation dependencies or cycles present.

There is a future cleanup opportunity to remove the now-constant
force_first and for_fini arguments from the _dl_sort_maps function.

Fixes commit 1df71d32fe5f5905ffd5d100e5e9ca8ad62 ("elf: Implement
force_first handling in _dl_sort_maps_dfs (bug 28937)").

Reviewed-by: DJ Delorie <dj@redhat.com>

1 files changed, 52 insertions, 100 deletions

diff --git a/elf/dl-fini.c b/elf/dl-fini.c
index 9acb64f47c..e201d36651 100644
--- a/elf/dl-fini.c
+++ b/elf/dl-fini.c
@@ -24,116 +24,68 @@
 void
 _dl_fini (void)
 {
-  /* Lots of fun ahead.  We have to call the destructors for all still
-     loaded objects, in all namespaces.  The problem is that the ELF
-     specification now demands that dependencies between the modules
-     are taken into account.  I.e., the destructor for a module is
-     called before the ones for any of its dependencies.
-
-     To make things more complicated, we cannot simply use the reverse
-     order of the constructors.  Since the user might have loaded objects
-     using `dlopen' there are possibly several other modules with its
-     dependencies to be taken into account.  Therefore we have to start
-     determining the order of the modules once again from the beginning.  */
-
-  /* We run the destructors of the main namespaces last.  As for the
-     other namespaces, we pick run the destructors in them in reverse
-     order of the namespace ID.  */
-#ifdef SHARED
-  int do_audit = 0;
- again:
-#endif
-  for (Lmid_t ns = GL(dl_nns) - 1; ns >= 0; --ns)
-    {
-      /* Protect against concurrent loads and unloads.  */
-      __rtld_lock_lock_recursive (GL(dl_load_lock));
-
-      unsigned int nloaded = GL(dl_ns)[ns]._ns_nloaded;
-      /* No need to do anything for empty namespaces or those used for
-	 auditing DSOs.  */
-      if (nloaded == 0
-#ifdef SHARED
-	  || GL(dl_ns)[ns]._ns_loaded->l_auditing != do_audit
-#endif
-	  )
-	__rtld_lock_unlock_recursive (GL(dl_load_lock));
-      else
-	{
+  /* Call destructors strictly in the reverse order of constructors.
+     This causes fewer surprises than some arbitrary reordering based
+     on new (relocation) dependencies.  None of the objects are
+     unmapped, so applications can deal with this if their DSOs remain
+     in a consistent state after destructors have run.  */
+
+  /* Protect against concurrent loads and unloads.  */
+  __rtld_lock_lock_recursive (GL(dl_load_lock));
+
+  /* Ignore objects which are opened during shutdown.  */
+  struct link_map *local_init_called_list = _dl_init_called_list;
+
+  for (struct link_map *l = local_init_called_list; l != NULL;
+       l = l->l_init_called_next)
+      /* Bump l_direct_opencount of all objects so that they
+	 are not dlclose()ed from underneath us.  */
+      ++l->l_direct_opencount;
+
+  /* After this point, everything linked from local_init_called_list
+     cannot be unloaded because of the reference counter update.  */
+  __rtld_lock_unlock_recursive (GL(dl_load_lock));
+
+  /* Perform two passes: One for non-audit modules, one for audit
+     modules.  This way, audit modules receive unload notifications
+     for non-audit objects, and the destructors for audit modules
+     still run.  */
 #ifdef SHARED
-	  _dl_audit_activity_nsid (ns, LA_ACT_DELETE);
+  int last_pass = GLRO(dl_naudit) > 0;
+  Lmid_t last_ns = -1;
+  for (int do_audit = 0; do_audit <= last_pass; ++do_audit)
 #endif
-
-	  /* Now we can allocate an array to hold all the pointers and
-	     copy the pointers in.  */
-	  struct link_map *maps[nloaded];
-
-	  unsigned int i;
-	  struct link_map *l;
-	  assert (nloaded != 0 || GL(dl_ns)[ns]._ns_loaded == NULL);
-	  for (l = GL(dl_ns)[ns]._ns_loaded, i = 0; l != NULL; l = l->l_next)
-	    /* Do not handle ld.so in secondary namespaces.  */
-	    if (l == l->l_real)
-	      {
-		assert (i < nloaded);
-
-		maps[i] = l;
-		l->l_idx = i;
-		++i;
-
-		/* Bump l_direct_opencount of all objects so that they
-		   are not dlclose()ed from underneath us.  */
-		++l->l_direct_opencount;
-	      }
-	  assert (ns != LM_ID_BASE || i == nloaded);
-	  assert (ns == LM_ID_BASE || i == nloaded || i == nloaded - 1);
-	  unsigned int nmaps = i;
-
-	  /* Now we have to do the sorting.  We can skip looking for the
-	     binary itself which is at the front of the search list for
-	     the main namespace.  */
-	  _dl_sort_maps (maps, nmaps, (ns == LM_ID_BASE), true);
-
-	  /* We do not rely on the linked list of loaded object anymore
-	     from this point on.  We have our own list here (maps).  The
-	     various members of this list cannot vanish since the open
-	     count is too high and will be decremented in this loop.  So
-	     we release the lock so that some code which might be called
-	     from a destructor can directly or indirectly access the
-	     lock.  */
-	  __rtld_lock_unlock_recursive (GL(dl_load_lock));
-
-	  /* 'maps' now contains the objects in the right order.  Now
-	     call the destructors.  We have to process this array from
-	     the front.  */
-	  for (i = 0; i < nmaps; ++i)
-	    {
-	      struct link_map *l = maps[i];
-
-	      if (l->l_init_called)
-		{
-		  _dl_call_fini (l);
+    for (struct link_map *l = local_init_called_list; l != NULL;
+	 l = l->l_init_called_next)
+      {
 #ifdef SHARED
-		  /* Auditing checkpoint: another object closed.  */
-		  _dl_audit_objclose (l);
+	if (GL(dl_ns)[l->l_ns]._ns_loaded->l_auditing != do_audit)
+	  continue;
+
+	/* Avoid back-to-back calls of _dl_audit_activity_nsid for the
+	   same namespace.  */
+	if (last_ns != l->l_ns)
+	  {
+	    if (last_ns >= 0)
+	      _dl_audit_activity_nsid (last_ns, LA_ACT_CONSISTENT);
+	    _dl_audit_activity_nsid (l->l_ns, LA_ACT_DELETE);
+	    last_ns = l->l_ns;
+	  }
 #endif
-		}
 
-	      /* Correct the previous increment.  */
-	      --l->l_direct_opencount;
-	    }
+	/* There is no need to re-enable exceptions because _dl_fini
+	   is not called from a context where exceptions are caught.  */
+	_dl_call_fini (l);
 
 #ifdef SHARED
-	  _dl_audit_activity_nsid (ns, LA_ACT_CONSISTENT);
+	/* Auditing checkpoint: another object closed.  */
+	_dl_audit_objclose (l);
 #endif
-	}
-    }
+      }
 
 #ifdef SHARED
-  if (! do_audit && GLRO(dl_naudit) > 0)
-    {
-      do_audit = 1;
-      goto again;
-    }
+  if (last_ns >= 0)
+    _dl_audit_activity_nsid (last_ns, LA_ACT_CONSISTENT);
 
   if (__glibc_unlikely (GLRO(dl_debug_mask) & DL_DEBUG_STATISTICS))
     _dl_debug_printf ("\nruntime linker statistics:\n"