* [PATCH] Fix handling of STB_GNU_UNIQUE in LD_TRACE_PRELINKING
@ 2010-03-23 10:34 Jakub Jelinek
2010-03-24 17:16 ` Ulrich Drepper
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2010-03-23 10:34 UTC (permalink / raw)
To: Ulrich Drepper; +Cc: Glibc hackers, Michael Matz
[-- Attachment #1: Type: text/plain, Size: 1556 bytes --]
Hi!
The STB_GNU_UNIQUE hash table breaks computation of conflicts
during LD_TRACE_PRELINKING, thus prelink doesn't add relocations into
.gnu.conflict section and programs might misbehave at runtime when
prelinked.
The problem is that to detect a conflict, _dl_debug_bindings in addition
to the normal lookup does also a lookup in undef_map->l_local_scope[0],
and if the lookup has different result, reports a conflict, otherwise
just non-conflicting lookup. But, as STB_GNU_UNIQUE hash table for
such symbols always gives the first lookup that entered the hash table,
conflicts are never reported for STB_GNU_UNIQUE lookups.
Attached are two possible patches. The second, shorter one, is all that is
needed currently (I believe). The difference between the patches is in
handling libraries with DT_SYMBOLIC and STB_GNU_UNIQUE symbols. Apparently
current GNU ld if there is a relocation against STB_GNU_UNIQUE symbol
defined in the library itself, doesn't emit any relocation and resolves at
link time (possibly through a relative relocation). Is that what we want
though? Then the STB_GNU_UNIQUE symbols aren't really unique...
If we want to change ld(1) behavior to always emit dynamic relocations
against STB_GNU_UNIQUE symbols defined in the library even
for -Wl,-Bsymbolic, then I'm afraid we need something like the former patch.
It can be pessimizing, reporting a conflict even when there is really none,
but I don't see how else to do this cheaply. And, -Wl,-Bsymbolic is very
rare I believe, so I don't care about that too much.
Jakub
[-- Attachment #2: S --]
[-- Type: text/plain, Size: 5469 bytes --]
2010-03-23 Jakub Jelinek <jakub@redhat.com>
* elf/dl-lookup.c (do_lookup_x): If tab->entries is NULL,
but tab->size != 0, just unlock and goto success, without
allocating anything or entering anything into the hash table.
(_dl_debug_bindings): Temporarily set tab->entries to NULL
around do_lookup_x in undef_map->l_local_scope[0]. If
undef_map->l_symbolic_in_local_scope, lookup also in
symbolic_searchlist of following libraries in l_local_scope
that have DT_SYMBOLIC set.
* elf/dl-deps.c (_dl_map_object_deps): Compute
l_symbolic_in_local_scope.
* include/link.h (struct link_map): Add l_symbolic_in_local_scope
bitfield.
--- libc/elf/dl-deps.c.jj 2009-05-16 19:23:29.000000000 +0200
+++ libc/elf/dl-deps.c 2010-03-23 09:24:57.000000000 +0100
@@ -1,5 +1,5 @@
/* Load the dependencies of a mapped object.
- Copyright (C) 1996-2003, 2004, 2005, 2006, 2007
+ Copyright (C) 1996-2003, 2004, 2005, 2006, 2007, 2010
Free Software Foundation, Inc.
This file is part of the GNU C Library.
@@ -554,7 +554,12 @@ Filters not supported with LD_TRACE_PREL
cnt = _dl_build_local_scope (l_initfini, l);
assert (cnt <= nlist);
for (j = 0; j < cnt; j++)
- l_initfini[j]->l_reserved = 0;
+ {
+ l_initfini[j]->l_reserved = 0;
+ if (j && __builtin_expect (l_initfini[j]->l_info[DT_SYMBOLIC]
+ != NULL, 0))
+ l->l_symbolic_in_local_scope = true;
+ }
l->l_local_scope[0] =
(struct r_scope_elem *) malloc (sizeof (struct r_scope_elem)
--- libc/elf/dl-lookup.c.jj 2009-12-16 00:10:37.000000000 +0100
+++ libc/elf/dl-lookup.c 2010-03-23 10:00:35.000000000 +0100
@@ -1,5 +1,6 @@
/* Look up a symbol in the loaded objects.
- Copyright (C) 1995-2005, 2006, 2007, 2009 Free Software Foundation, Inc.
+ Copyright (C) 1995-2005, 2006, 2007, 2009, 2010
+ Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
@@ -414,6 +415,20 @@ do_lookup_x (const char *undef_name, uin
assert (!RTLD_CHECK_FOREIGN_CALL);
#endif
+#ifdef SHARED
+ /* If tab->entries is NULL, but tab->size is not, it means
+ this is the second, conflict finding, lookup for
+ LD_TRACE_PRELINKING in _dl_debug_bindings. Don't
+ allocate anything and don't enter anything into the
+ hash table. */
+ if (__builtin_expect (tab->size, 0))
+ {
+ assert (GLRO(dl_debug_mask) & DL_DEBUG_PRELINK);
+ __rtld_lock_unlock_recursive (tab->lock);
+ goto success;
+ }
+#endif
+
#define INITIAL_NUNIQUE_SYM_TABLE 31
size = INITIAL_NUNIQUE_SYM_TABLE;
entries = calloc (sizeof (struct unique_sym), size);
@@ -917,13 +932,48 @@ _dl_debug_bindings (const char *undef_na
{
const uint_fast32_t new_hash = dl_new_hash (undef_name);
unsigned long int old_hash = 0xffffffff;
+ struct unique_sym *saved_entries
+ = GL(dl_ns)[LM_ID_BASE]._ns_unique_sym_table.entries;
+ GL(dl_ns)[LM_ID_BASE]._ns_unique_sym_table.entries = NULL;
do_lookup_x (undef_name, new_hash, &old_hash, *ref, &val,
undef_map->l_local_scope[0], 0, version, 0, NULL,
type_class, undef_map);
-
if (val.s != value->s || val.m != value->m)
conflict = 1;
+ else if (__builtin_expect (undef_map->l_symbolic_in_local_scope, 0)
+ && val.s
+ && __builtin_expect (ELFW(ST_BIND) (val.s->st_info),
+ STB_GLOBAL) == STB_GNU_UNIQUE)
+ {
+ /* If it is STB_GNU_UNIQUE and undef_map's l_local_scope
+ contains any DT_SYMBOLIC libraries, unfortunately there
+ can be conflicts even if the above is equal. As symbol
+ resolution goes from the last library to the first and
+ if a STB_GNU_UNIQUE symbol is found in some late DT_SYMBOLIC
+ library, it would be the one that is looked up. */
+ struct sym_val val2 = { NULL, NULL };
+ size_t n;
+ struct r_scope_elem *scope = undef_map->l_local_scope[0];
+
+ for (n = 0; n < scope->r_nlist; n++)
+ if (scope->r_list[n] == val.m)
+ break;
+
+ for (n++; n < scope->r_nlist; n++)
+ if (scope->r_list[n]->l_info[DT_SYMBOLIC] != NULL
+ && do_lookup_x (undef_name, new_hash, &old_hash, *ref,
+ &val2,
+ &scope->r_list[n]->l_symbolic_searchlist,
+ 0, version, 0, NULL, type_class,
+ undef_map) > 0)
+ {
+ conflict = 1;
+ val = val2;
+ break;
+ }
+ }
+ GL(dl_ns)[LM_ID_BASE]._ns_unique_sym_table.entries = saved_entries;
}
if (value->s)
--- libc/include/link.h.jj 2009-12-16 00:10:37.000000000 +0100
+++ libc/include/link.h 2010-03-23 09:27:18.000000000 +0100
@@ -1,6 +1,6 @@
/* Data structure for communication from the run-time dynamic linker for
loaded ELF shared objects.
- Copyright (C) 1995-2006, 2007, 2009 Free Software Foundation, Inc.
+ Copyright (C) 1995-2006, 2007, 2009, 2010 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
@@ -188,6 +188,10 @@ struct link_map
unsigned int l_contiguous:1; /* Nonzero if inter-segment holes are
mprotected or if no holes are present at
all. */
+ unsigned int l_symbolic_in_local_scope:1; /* Nonzero if l_local_scope
+ during LD_TRACE_PRELINKING=1
+ contains any DT_SYMBOLIC
+ libraries. */
/* Collected information about own RPATH directories. */
struct r_search_path_struct l_rpath_dirs;
[-- Attachment #3: S2 --]
[-- Type: text/plain, Size: 2194 bytes --]
2010-03-23 Jakub Jelinek <jakub@redhat.com>
* elf/dl-lookup.c (do_lookup_x): If tab->entries is NULL,
but tab->size != 0, just unlock and goto success, without
allocating anything or entering anything into the hash table.
(_dl_debug_bindings): Temporarily set tab->entries to NULL
around do_lookup_x in undef_map->l_local_scope[0].
--- libc/elf/dl-lookup.c.jj 2009-12-16 00:10:37.000000000 +0100
+++ libc/elf/dl-lookup.c 2010-03-23 11:03:45.850560036 +0100
@@ -1,5 +1,6 @@
/* Look up a symbol in the loaded objects.
- Copyright (C) 1995-2005, 2006, 2007, 2009 Free Software Foundation, Inc.
+ Copyright (C) 1995-2005, 2006, 2007, 2009, 2010
+ Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
@@ -414,6 +415,20 @@ do_lookup_x (const char *undef_name, uin
assert (!RTLD_CHECK_FOREIGN_CALL);
#endif
+#ifdef SHARED
+ /* If tab->entries is NULL, but tab->size is not, it means
+ this is the second, conflict finding, lookup for
+ LD_TRACE_PRELINKING in _dl_debug_bindings. Don't
+ allocate anything and don't enter anything into the
+ hash table. */
+ if (__builtin_expect (tab->size, 0))
+ {
+ assert (GLRO(dl_debug_mask) & DL_DEBUG_PRELINK);
+ __rtld_lock_unlock_recursive (tab->lock);
+ goto success;
+ }
+#endif
+
#define INITIAL_NUNIQUE_SYM_TABLE 31
size = INITIAL_NUNIQUE_SYM_TABLE;
entries = calloc (sizeof (struct unique_sym), size);
@@ -917,13 +932,17 @@ _dl_debug_bindings (const char *undef_na
{
const uint_fast32_t new_hash = dl_new_hash (undef_name);
unsigned long int old_hash = 0xffffffff;
+ struct unique_sym *saved_entries
+ = GL(dl_ns)[LM_ID_BASE]._ns_unique_sym_table.entries;
+ GL(dl_ns)[LM_ID_BASE]._ns_unique_sym_table.entries = NULL;
do_lookup_x (undef_name, new_hash, &old_hash, *ref, &val,
undef_map->l_local_scope[0], 0, version, 0, NULL,
type_class, undef_map);
if (val.s != value->s || val.m != value->m)
conflict = 1;
+ GL(dl_ns)[LM_ID_BASE]._ns_unique_sym_table.entries = saved_entries;
}
if (value->s)
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Fix handling of STB_GNU_UNIQUE in LD_TRACE_PRELINKING
2010-03-23 10:34 [PATCH] Fix handling of STB_GNU_UNIQUE in LD_TRACE_PRELINKING Jakub Jelinek
@ 2010-03-24 17:16 ` Ulrich Drepper
0 siblings, 0 replies; 2+ messages in thread
From: Ulrich Drepper @ 2010-03-24 17:16 UTC (permalink / raw)
To: libc-hacker
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
On 03/23/2010 03:33 AM, Jakub Jelinek wrote:
> Attached are two possible patches. The second, shorter one, is all that is
> needed currently (I believe). The difference between the patches is in
> handling libraries with DT_SYMBOLIC and STB_GNU_UNIQUE symbols. Apparently
> current GNU ld if there is a relocation against STB_GNU_UNIQUE symbol
> defined in the library itself, doesn't emit any relocation and resolves at
> link time (possibly through a relative relocation). Is that what we want
> though? Then the STB_GNU_UNIQUE symbols aren't really unique...
Right. The linker has to be fixed. I've applied the first patch.
- --
â§ Ulrich Drepper â§ Red Hat, Inc. â§ 444 Castro St â§ Mountain View, CA â
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/
iEYEARECAAYFAkuqSNcACgkQ2ijCOnn/RHQkNACgu14Wmhn6JxpmuDbE3PEUG82a
qO8AniThFZts0yE1zJ+IXaBQjD3P1XOW
=2rjh
-----END PGP SIGNATURE-----
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2010-03-24 17:16 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2010-03-23 10:34 [PATCH] Fix handling of STB_GNU_UNIQUE in LD_TRACE_PRELINKING Jakub Jelinek
2010-03-24 17:16 ` Ulrich Drepper
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).