about summary refs log tree commit diff
path: root/elf/dl-sort-maps.c
diff options
context:
space:
mode:
authorChung-Lin Tang <cltang@codesourcery.com>2021-10-21 21:41:22 +0800
committerAdhemerval Zanella <adhemerval.zanella@linaro.org>2021-10-21 11:23:53 -0300
commit15a0c5730d1d5aeb95f50c9ec7470640084feae8 (patch)
tree3c38108344adce9db163005cb197b4300a6fff94 /elf/dl-sort-maps.c
parente6fd79f3795d46dfb583e124be49fc063bc3d58b (diff)
downloadglibc-15a0c5730d1d5aeb95f50c9ec7470640084feae8.tar.gz
glibc-15a0c5730d1d5aeb95f50c9ec7470640084feae8.tar.xz
glibc-15a0c5730d1d5aeb95f50c9ec7470640084feae8.zip
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>
Diffstat (limited to 'elf/dl-sort-maps.c')
-rw-r--r--elf/dl-sort-maps.c208
1 files changed, 194 insertions, 14 deletions
diff --git a/elf/dl-sort-maps.c b/elf/dl-sort-maps.c
index d21770267a..a274ed66cc 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.  */