elf: Fix slow DSO sorting behavior in dynamic loader (BZ #17645)

This second patch contains the actual implementation of a new sorting algorithm
for shared objects in the dynamic loader, which solves the slow behavior that
the current "old" algorithm falls into when the DSO set contains circular
dependencies.

The new algorithm implemented here is simply depth-first search (DFS) to obtain
the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1
bitfield is added to struct link_map to more elegantly facilitate such a search.

The DFS algorithm is applied to the input maps[nmap-1] backwards towards
maps[0]. This has the effect of a more "shallow" recursion depth in general
since the input is in BFS. Also, when combined with the natural order of
processing l_initfini[] at each node, this creates a resulting output sorting
closer to the intuitive "left-to-right" order in most cases.

Another notable implementation adjustment related to this _dl_sort_maps change
is the removing of two char arrays 'used' and 'done' in _dl_close_worker to
represent two per-map attributes. This has been changed to simply use two new
bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows
discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes
do along the way.

Tunable support for switching between different sorting algorithms at runtime is
also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1
(old algorithm) and 2 (new DFS algorithm) has been added. At time of commit
of this patch, the default setting is 1 (old algorithm).

Signed-off-by: Chung-Lin Tang  <cltang@codesourcery.com>
Reviewed-by: Adhemerval Zanella  <adhemerval.zanella@linaro.org>

10 files changed, 231 insertions, 41 deletions

diff --git a/elf/dl-close.c b/elf/dl-close.c
index cfe0f1c..4f5cfcc 100644
--- a/elf/dl-close.c
+++ b/elf/dl-close.c
@@ -167,8 +167,6 @@ _dl_close_worker (struct link_map *map, bool force)
 
   bool any_tls = false;
   const unsigned int nloaded = ns->_ns_nloaded;
-  char used[nloaded];
-  char done[nloaded];
   struct link_map *maps[nloaded];
 
   /* Run over the list and assign indexes to the link maps and enter
@@ -176,24 +174,21 @@ _dl_close_worker (struct link_map *map, bool force)
   int idx = 0;
   for (struct link_map *l = ns->_ns_loaded; l != NULL; l = l->l_next)
     {
+      l->l_map_used = 0;
+      l->l_map_done = 0;
       l->l_idx = idx;
       maps[idx] = l;
       ++idx;
-
     }
   assert (idx == nloaded);
 
-  /* Prepare the bitmaps.  */
-  memset (used, '\0', sizeof (used));
-  memset (done, '\0', sizeof (done));
-
   /* Keep track of the lowest index link map we have covered already.  */
   int done_index = -1;
   while (++done_index < nloaded)
     {
       struct link_map *l = maps[done_index];
 
-      if (done[done_index])
+      if (l->l_map_done)
 	/* Already handled.  */
 	continue;
 
@@ -204,12 +199,12 @@ _dl_close_worker (struct link_map *map, bool force)
 	  /* See CONCURRENCY NOTES in cxa_thread_atexit_impl.c to know why
 	     acquire is sufficient and correct.  */
 	  && atomic_load_acquire (&l->l_tls_dtor_count) == 0
-	  && !used[done_index])
+	  && !l->l_map_used)
 	continue;
 
       /* We need this object and we handle it now.  */
-      done[done_index] = 1;
-      used[done_index] = 1;
+      l->l_map_used = 1;
+      l->l_map_done = 1;
       /* Signal the object is still needed.  */
       l->l_idx = IDX_STILL_USED;
 
@@ -225,9 +220,9 @@ _dl_close_worker (struct link_map *map, bool force)
 		{
 		  assert ((*lp)->l_idx >= 0 && (*lp)->l_idx < nloaded);
 
-		  if (!used[(*lp)->l_idx])
+		  if (!(*lp)->l_map_used)
 		    {
-		      used[(*lp)->l_idx] = 1;
+		      (*lp)->l_map_used = 1;
 		      /* If we marked a new object as used, and we've
 			 already processed it, then we need to go back
 			 and process again from that point forward to
@@ -250,9 +245,9 @@ _dl_close_worker (struct link_map *map, bool force)
 	      {
 		assert (jmap->l_idx >= 0 && jmap->l_idx < nloaded);
 
-		if (!used[jmap->l_idx])
+		if (!jmap->l_map_used)
 		  {
-		    used[jmap->l_idx] = 1;
+		    jmap->l_map_used = 1;
 		    if (jmap->l_idx - 1 < done_index)
 		      done_index = jmap->l_idx - 1;
 		  }
@@ -262,8 +257,7 @@ _dl_close_worker (struct link_map *map, bool force)
 
   /* Sort the entries.  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 + (nsid == LM_ID_BASE), nloaded - (nsid == LM_ID_BASE),
-		 used + (nsid == LM_ID_BASE), true);
+  _dl_sort_maps (maps, nloaded, (nsid == LM_ID_BASE), true);
 
   /* Call all termination functions at once.  */
 #ifdef SHARED
@@ -280,7 +274,7 @@ _dl_close_worker (struct link_map *map, bool force)
       /* All elements must be in the same namespace.  */
       assert (imap->l_ns == nsid);
 
-      if (!used[i])
+      if (!imap->l_map_used)
 	{
 	  assert (imap->l_type == lt_loaded && !imap->l_nodelete_active);
 
@@ -333,7 +327,7 @@ _dl_close_worker (struct link_map *map, bool force)
 	  if (i < first_loaded)
 	    first_loaded = i;
 	}
-      /* Else used[i].  */
+      /* Else imap->l_map_used.  */
       else if (imap->l_type == lt_loaded)
 	{
 	  struct r_scope_elem *new_list = NULL;
@@ -560,7 +554,7 @@ _dl_close_worker (struct link_map *map, bool force)
   for (unsigned int i = first_loaded; i < nloaded; ++i)
     {
       struct link_map *imap = maps[i];
-      if (!used[i])
+      if (!imap->l_map_used)
 	{
 	  assert (imap->l_type == lt_loaded);
 
diff --git a/elf/dl-deps.c b/elf/dl-deps.c
index 087a49b..237d963 100644
--- a/elf/dl-deps.c
+++ b/elf/dl-deps.c
@@ -613,10 +613,9 @@ Filters not supported with LD_TRACE_PRELINKING"));
 
   /* If libc.so.6 is the main map, it participates in the sort, so
      that the relocation order is correct regarding libc.so.6.  */
-  if (l_initfini[0] == GL (dl_ns)[l_initfini[0]->l_ns].libc_map)
-    _dl_sort_maps (l_initfini, nlist, NULL, false);
-  else
-    _dl_sort_maps (&l_initfini[1], nlist - 1, NULL, false);
+  _dl_sort_maps (l_initfini, nlist,
+		 (l_initfini[0] != GL (dl_ns)[l_initfini[0]->l_ns].libc_map),
+		 false);
 
   /* Terminate the list of dependencies.  */
   l_initfini[nlist] = NULL;
diff --git a/elf/dl-fini.c b/elf/dl-fini.c
index 6dbdfe4..c683884 100644
--- a/elf/dl-fini.c
+++ b/elf/dl-fini.c
@@ -92,8 +92,7 @@ _dl_fini (void)
 	  /* 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 + (ns == LM_ID_BASE), nmaps - (ns == LM_ID_BASE),
-			 NULL, true);
+	  _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
diff --git a/elf/dl-sort-maps.c b/elf/dl-sort-maps.c
index d217702..a274ed6 100644
--- a/elf/dl-sort-maps.c
+++ b/elf/dl-sort-maps.c
@@ -16,16 +16,24 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
+#include <assert.h>
 #include <ldsodefs.h>
+#include <elf/dl-tunables.h>
 
+/* Note: this is the older, "original" sorting algorithm, being used as
+   default up to 2.35.
 
-/* Sort array MAPS according to dependencies of the contained objects.
-   Array USED, if non-NULL, is permutated along MAPS.  If FOR_FINI this is
-   called for finishing an object.  */
-void
-_dl_sort_maps (struct link_map **maps, unsigned int nmaps, char *used,
-	       bool for_fini)
+   Sort array MAPS according to dependencies of the contained objects.
+   If FOR_FINI is true, this is called for finishing an object.  */
+static void
+_dl_sort_maps_original (struct link_map **maps, unsigned int nmaps,
+			unsigned int skip, bool for_fini)
 {
+  /* Allows caller to do the common optimization of skipping the first map,
+     usually the main binary.  */
+  maps += skip;
+  nmaps -= skip;
+
   /* A list of one element need not be sorted.  */
   if (nmaps <= 1)
     return;
@@ -66,14 +74,6 @@ _dl_sort_maps (struct link_map **maps, unsigned int nmaps, char *used,
 			   (k - i) * sizeof (maps[0]));
 		  maps[k] = thisp;
 
-		  if (used != NULL)
-		    {
-		      char here_used = used[i];
-		      memmove (&used[i], &used[i + 1],
-			       (k - i) * sizeof (used[0]));
-		      used[k] = here_used;
-		    }
-
 		  if (seen[i + 1] > nmaps - i)
 		    {
 		      ++i;
@@ -120,3 +120,183 @@ _dl_sort_maps (struct link_map **maps, unsigned int nmaps, char *used,
     next:;
     }
 }
+
+#if !HAVE_TUNABLES
+/* In this case, just default to the original algorithm.  */
+strong_alias (_dl_sort_maps_original, _dl_sort_maps);
+#else
+
+/* We use a recursive function due to its better clarity and ease of
+   implementation, as well as faster execution speed. We already use
+   alloca() for list allocation during the breadth-first search of
+   dependencies in _dl_map_object_deps(), and this should be on the
+   same order of worst-case stack usage.
+
+   Note: the '*rpo' parameter is supposed to point to one past the
+   last element of the array where we save the sort results, and is
+   decremented before storing the current map at each level.  */
+
+static void
+dfs_traversal (struct link_map ***rpo, struct link_map *map,
+	       bool *do_reldeps)
+{
+  if (map->l_visited)
+    return;
+
+  map->l_visited = 1;
+
+  if (map->l_initfini)
+    {
+      for (int i = 0; map->l_initfini[i] != NULL; i++)
+	{
+	  struct link_map *dep = map->l_initfini[i];
+	  if (dep->l_visited == 0
+	      && dep->l_main_map == 0)
+	    dfs_traversal (rpo, dep, do_reldeps);
+	}
+    }
+
+  if (__glibc_unlikely (do_reldeps != NULL && map->l_reldeps != NULL))
+    {
+      /* Indicate that we encountered relocation dependencies during
+	 traversal.  */
+      *do_reldeps = true;
+
+      for (int m = map->l_reldeps->act - 1; m >= 0; m--)
+	{
+	  struct link_map *dep = map->l_reldeps->list[m];
+	  if (dep->l_visited == 0
+	      && dep->l_main_map == 0)
+	    dfs_traversal (rpo, dep, do_reldeps);
+	}
+    }
+
+  *rpo -= 1;
+  **rpo = map;
+}
+
+/* Topologically sort array MAPS according to dependencies of the contained
+   objects.  */
+
+static void
+_dl_sort_maps_dfs (struct link_map **maps, unsigned int nmaps,
+		   unsigned int skip __attribute__ ((unused)), bool for_fini)
+{
+  for (int i = nmaps - 1; i >= 0; i--)
+    maps[i]->l_visited = 0;
+
+  /* We apply DFS traversal for each of maps[i] until the whole total order
+     is found and we're at the start of the Reverse-Postorder (RPO) sequence,
+     which is a topological sort.
+
+     We go from maps[nmaps - 1] backwards towards maps[0] at this level.
+     Due to the breadth-first search (BFS) ordering we receive, going
+     backwards usually gives a more shallow depth-first recursion depth,
+     adding more stack usage safety. Also, combined with the natural
+     processing order of l_initfini[] at each node during DFS, this maintains
+     an ordering closer to the original link ordering in the sorting results
+     under most simpler cases.
+
+     Another reason we order the top level backwards, it that maps[0] is
+     usually exactly the main object of which we're in the midst of
+     _dl_map_object_deps() processing, and maps[0]->l_initfini[] is still
+     blank. If we start the traversal from maps[0], since having no
+     dependencies yet filled in, maps[0] will always be immediately
+     incorrectly placed at the last place in the order (first in reverse).
+     Adjusting the order so that maps[0] is last traversed naturally avoids
+     this problem.
+
+     Further, the old "optimization" of skipping the main object at maps[0]
+     from the call-site (i.e. _dl_sort_maps(maps+1,nmaps-1)) is in general
+     no longer valid, since traversing along object dependency-links
+     may "find" the main object even when it is not included in the initial
+     order (e.g. a dlopen()'ed shared object can have circular dependencies
+     linked back to itself). In such a case, traversing N-1 objects will
+     create a N-object result, and raise problems.
+
+     To summarize, just passing in the full list, and iterating from back
+     to front makes things much more straightforward.  */
+
+  /* Array to hold RPO sorting results, before we copy back to maps[].  */
+  struct link_map *rpo[nmaps];
+
+  /* The 'head' position during each DFS iteration. Note that we start at
+     one past the last element due to first-decrement-then-store (see the
+     bottom of above dfs_traversal() routine).  */
+  struct link_map **rpo_head = &rpo[nmaps];
+
+  bool do_reldeps = false;
+  bool *do_reldeps_ref = (for_fini ? &do_reldeps : NULL);
+
+  for (int i = nmaps - 1; i >= 0; i--)
+    {
+      dfs_traversal (&rpo_head, maps[i], do_reldeps_ref);
+
+      /* We can break early if all objects are already placed.  */
+      if (rpo_head == rpo)
+	goto end;
+    }
+  assert (rpo_head == rpo);
+
+ end:
+  /* Here we may do a second pass of sorting, using only l_initfini[]
+     static dependency links. This is avoided if !FOR_FINI or if we didn't
+     find any reldeps in the first DFS traversal.
+
+     The reason we do this is: while it is unspecified how circular
+     dependencies should be handled, the presumed reasonable behavior is to
+     have destructors to respect static dependency links as much as possible,
+     overriding reldeps if needed. And the first sorting pass, which takes
+     l_initfini/l_reldeps links equally, may not preserve this priority.
+
+     Hence we do a 2nd sorting pass, taking only DT_NEEDED links into account
+     (see how the do_reldeps argument to dfs_traversal() is NULL below).  */
+  if (do_reldeps)
+    {
+      for (int i = nmaps - 1; i >= 0; i--)
+	rpo[i]->l_visited = 0;
+
+      struct link_map **maps_head = &maps[nmaps];
+      for (int i = nmaps - 1; i >= 0; i--)
+	{
+	  dfs_traversal (&maps_head, rpo[i], NULL);
+
+	  /* We can break early if all objects are already placed.
+	     The below memcpy is not needed in the do_reldeps case here,
+	     since we wrote back to maps[] during DFS traversal.  */
+	  if (maps_head == maps)
+	    return;
+	}
+      assert (maps_head == maps);
+      return;
+    }
+
+  memcpy (maps, rpo, sizeof (struct link_map *) * nmaps);
+}
+
+void
+_dl_sort_maps_init (void)
+{
+  int32_t algorithm = TUNABLE_GET (glibc, rtld, dynamic_sort, int32_t, NULL);
+  GLRO(dl_dso_sort_algo) = algorithm == 1 ? dso_sort_algorithm_original
+					  : dso_sort_algorithm_dfs;
+}
+
+void
+_dl_sort_maps (struct link_map **maps, unsigned int nmaps,
+	       unsigned int skip, bool for_fini)
+{
+  /* It can be tempting to use a static function pointer to store and call
+     the current selected sorting algorithm routine, but experimentation
+     shows that current processors still do not handle indirect branches
+     that efficiently, plus a static function pointer will involve
+     PTR_MANGLE/DEMANGLE, further impairing performance of small, common
+     input cases. A simple if-case with direct function calls appears to
+     be the fastest.  */
+  if (__glibc_likely (GLRO(dl_dso_sort_algo) == dso_sort_algorithm_original))
+    _dl_sort_maps_original (maps, nmaps, skip, for_fini);
+  else
+    _dl_sort_maps_dfs (maps, nmaps, skip, for_fini);
+}
+
+#endif /* HAVE_TUNABLES.  */
diff --git a/elf/dl-support.c b/elf/dl-support.c
index d99c1f1..98d5d8d 100644
--- a/elf/dl-support.c
+++ b/elf/dl-support.c
@@ -166,6 +166,8 @@ size_t _dl_phnum;
 uint64_t _dl_hwcap;
 uint64_t _dl_hwcap2;
 
+enum dso_sort_algorithm _dl_dso_sort_algo;
+
 /* The value of the FPU control word the kernel will preset in hardware.  */
 fpu_control_t _dl_fpu_control = _FPU_DEFAULT;
 
diff --git a/elf/dl-sysdep.c b/elf/dl-sysdep.c
index 2c684c2..4dc366e 100644
--- a/elf/dl-sysdep.c
+++ b/elf/dl-sysdep.c
@@ -231,6 +231,9 @@ _dl_sysdep_start (void **start_argptr,
 
   __tunables_init (_environ);
 
+  /* Initialize DSO sorting algorithm after tunables.  */
+  _dl_sort_maps_init ();
+
 #ifdef DL_SYSDEP_INIT
   DL_SYSDEP_INIT;
 #endif
diff --git a/elf/dl-tunables.list b/elf/dl-tunables.list
index 8ddd4a2..46ffb23 100644
--- a/elf/dl-tunables.list
+++ b/elf/dl-tunables.list
@@ -156,4 +156,13 @@ glibc {
       security_level: SXID_IGNORE
     }
   }
+
+  rtld {
+    dynamic_sort {
+      type: INT_32
+      minval: 1
+      maxval: 2
+      default: 1
+    }
+  }
 }
diff --git a/elf/dso-sort-tests-1.def b/elf/dso-sort-tests-1.def
index 873ddf5..5f7f18e 100644
--- a/elf/dso-sort-tests-1.def
+++ b/elf/dso-sort-tests-1.def
@@ -62,5 +62,5 @@ output: b>a>{}<a<b
 # The below expected outputs are what the two algorithms currently produce
 # respectively, for regression testing purposes.
 tst-bz15311: {+a;+e;+f;+g;+d;%d;-d;-g;-f;-e;-a};a->b->c->d;d=>[ba];c=>a;b=>e=>a;c=>f=>b;d=>g=>c
-xfail_output(glibc.rtld.dynamic_sort=1): {+a[d>c>b>a>];+e[e>];+f[f>];+g[g>];+d[];%d(b(e(a()))a()g(c(a()f(b(e(a()))))));-d[];-g[];-f[];-e[];-a[<a<c<d<g<f<b<e];}
+output(glibc.rtld.dynamic_sort=1): {+a[d>c>b>a>];+e[e>];+f[f>];+g[g>];+d[];%d(b(e(a()))a()g(c(a()f(b(e(a()))))));-d[];-g[];-f[];-e[];-a[<a<c<d<g<f<b<e];}
 output(glibc.rtld.dynamic_sort=2): {+a[d>c>b>a>];+e[e>];+f[f>];+g[g>];+d[];%d(b(e(a()))a()g(c(a()f(b(e(a()))))));-d[];-g[];-f[];-e[];-a[<g<f<a<b<c<d<e];}
diff --git a/elf/rtld.c b/elf/rtld.c
index b8ba2d8..be2d5d8 100644
--- a/elf/rtld.c
+++ b/elf/rtld.c
@@ -1391,6 +1391,9 @@ dl_main (const ElfW(Phdr) *phdr,
       main_map->l_name = (char *) "";
       *user_entry = main_map->l_entry;
 
+      /* Set bit indicating this is the main program map.  */
+      main_map->l_main_map = 1;
+
 #ifdef HAVE_AUX_VECTOR
       /* Adjust the on-stack auxiliary vector so that it looks like the
 	 binary was executed directly.  */
diff --git a/elf/tst-rtld-list-tunables.exp b/elf/tst-rtld-list-tunables.exp
index 9f66c52..9bf5727 100644
--- a/elf/tst-rtld-list-tunables.exp
+++ b/elf/tst-rtld-list-tunables.exp
@@ -10,5 +10,6 @@ glibc.malloc.tcache_max: 0x0 (min: 0x0, max: 0x[f]+)
 glibc.malloc.tcache_unsorted_limit: 0x0 (min: 0x0, max: 0x[f]+)
 glibc.malloc.top_pad: 0x0 (min: 0x0, max: 0x[f]+)
 glibc.malloc.trim_threshold: 0x0 (min: 0x0, max: 0x[f]+)
+glibc.rtld.dynamic_sort: 1 (min: 1, max: 2)
 glibc.rtld.nns: 0x4 (min: 0x1, max: 0x10)
 glibc.rtld.optional_static_tls: 0x200 (min: 0x0, max: 0x[f]+)