* sysdeps/generic/ldsodefs.h (struct rtld_global): Move all [USE_TLS]

	members to the end, so a libpthread compiled with !USE_TLS will still
	find other members properly.

	* sysdeps/i386/i486/bits/string.h (__strcpy_g): Add dummy output
	operand for DEST memory.  Fix dummy input operand to use SRC.
	Reported by Davin McCall <davmac@ozonline.com.au>.

	* sysdeps/generic/libc-tls.c (__libc_setup_tls): Account for TCB
	alignment when initializing the DTV entry.

	* elf/dl-load.c (_dl_map_object_from_fd): If we hit a TLS segment
	when TLS has not been set up, try to set it up if we can.
	* elf/tst-tls4.c: Revert last change.
	* elf/tst-tls5.c: Likewise.
	* elf/tst-tls6.c: Likewise.
	* elf/tst-tls7.c: Likewise.
	* elf/tst-tls8.c: Likewise.
	* elf/tst-tls9.c: Likewise.

	* sysdeps/generic/dl-tls.c [SHARED] (_dl_tls_setup): New function.
	* sysdeps/generic/ldsodefs.h: Declare it.
	* elf/Versions (ld: GLIBC_PRIVATE): Add it.
	* sysdeps/generic/libc-tls.c (init_slotinfo): New static inline
	function, broken out of __libc_setup_tls.
	(init_static_tls): Likewise.
	(__libc_setup_tls): Call them.
	(_dl_tls_setup): New function, uses new subroutines.

	* elf/dl-close.c (free_slotinfo): Make argument pointer to pointer.
	Clear the pointer when returning true.
	(libc_freeres_fn) [SHARED]: If GL(dl_initial_dtv) is null, free the
	first element of the slotinfo list too.

	* sysdeps/generic/dl-tls.c (_dl_determine_tlsoffset): Define only if
	[SHARED].

	* sysdeps/generic/ldsodefs.h (_dl_next_tls_modid): Declare as hidden.
	(_dl_determine_tlsoffset): Likewise.

	* elf/rtld.c (_dl_initial_error_catch_tsd): Renamed from
	startup_error_tsd, made global.
	(dl_main): Update initialization.
	* elf/dl-tsd.c: Likewise.
	* sysdeps/generic/ldsodefs.h: Declare it.

3 files changed, 96 insertions, 30 deletions

diff --git a/sysdeps/generic/dl-tls.c b/sysdeps/generic/dl-tls.c
index 850df8d76f..550de56530 100644
--- a/sysdeps/generic/dl-tls.c
+++ b/sysdeps/generic/dl-tls.c
@@ -110,6 +110,7 @@ _dl_next_tls_modid (void)
   return result;
 }
 
+# ifdef SHARED
 
 void
 internal_function
@@ -204,6 +205,41 @@ _dl_determine_tlsoffset (void)
 }
 
 
+/* This is called only when the data structure setup was skipped at startup,
+   when there was no need for it then.  Now we have dynamically loaded
+   something needing TLS, or libpthread needs it.  */
+int
+internal_function
+_dl_tls_setup (void)
+{
+  assert (GL(dl_tls_dtv_slotinfo_list) == NULL);
+  assert (GL(dl_tls_max_dtv_idx) == 0);
+
+  const size_t nelem = 2 + TLS_SLOTINFO_SURPLUS;
+
+  GL(dl_tls_dtv_slotinfo_list) =
+    malloc (sizeof (struct dtv_slotinfo_list)
+	    + nelem * sizeof (struct dtv_slotinfo));
+  if (GL(dl_tls_dtv_slotinfo_list) == NULL)
+    return -1;
+
+  memset (GL(dl_tls_dtv_slotinfo_list)->slotinfo, '\0',
+	  nelem * sizeof (struct dtv_slotinfo));
+  GL(dl_tls_dtv_slotinfo_list)->len = nelem;
+  GL(dl_tls_dtv_slotinfo_list)->next = NULL;
+
+  /* Number of elements in the static TLS block.  It can't be zero
+     because of various assumptions.  The one element is null.  */
+  GL(dl_tls_static_nelem) = GL(dl_tls_max_dtv_idx) = 1;
+
+  /* This initializes more variables for us.  */
+  _dl_determine_tlsoffset ();
+
+  return 0;
+}
+rtld_hidden_def (_dl_tls_setup)
+# endif
+
 static void *
 internal_function
 allocate_dtv (void *result)
diff --git a/sysdeps/generic/libc-tls.c b/sysdeps/generic/libc-tls.c
index a6a0504f44..756a2bf57c 100644
--- a/sysdeps/generic/libc-tls.c
+++ b/sysdeps/generic/libc-tls.c
@@ -74,6 +74,35 @@ size_t _dl_tls_generation;
 TLS_INIT_HELPER
 #endif
 
+static inline void
+init_slotinfo (void)
+{
+  /* Create the slotinfo list.  */
+  static_slotinfo.si.len = (((char *) (&static_slotinfo + 1)
+			     - (char *) &static_slotinfo.si.slotinfo[0])
+			    / sizeof static_slotinfo.si.slotinfo[0]);
+  // static_slotinfo.si.next = NULL;	already zero
+
+  /* The slotinfo list.  Will be extended by the code doing dynamic
+     linking.  */
+  GL(dl_tls_max_dtv_idx) = 1;
+  GL(dl_tls_dtv_slotinfo_list) = &static_slotinfo.si;
+}
+
+static inline void
+init_static_tls (size_t memsz, size_t align)
+{
+  /* That is the size of the TLS memory for this object.  The initialized
+     value of _dl_tls_static_size is provided by dl-open.c to request some
+     surplus that permits dynamic loading of modules with IE-model TLS.  */
+  GL(dl_tls_static_size) = roundup (memsz + GL(dl_tls_static_size),
+				    TLS_TCB_ALIGN);
+  GL(dl_tls_static_used) = memsz;
+  /* The alignment requirement for the static TLS block.  */
+  GL(dl_tls_static_align) = align;
+  /* Number of elements in the static TLS block.  */
+  GL(dl_tls_static_nelem) = GL(dl_tls_max_dtv_idx);
+}
 
 void
 __libc_setup_tls (size_t tcbsize, size_t tcbalign)
@@ -117,8 +146,8 @@ __libc_setup_tls (size_t tcbsize, size_t tcbalign)
      to request some surplus that permits dynamic loading of modules with
      IE-model TLS.  */
 # if TLS_TCB_AT_TP
-  tlsblock = __sbrk (roundup (memsz, tcbalign) + tcbsize + max_align
-		     + GL(dl_tls_static_size));
+  tcb_offset = roundup (memsz + GL(dl_tls_static_size), tcbalign);
+  tlsblock = __sbrk (tcb_offset + tcbsize + max_align);
 # elif TLS_DTV_AT_TP
   tlsblock = __sbrk (roundup (tcbsize, align) + memsz + max_align
 		     + GL(dl_tls_static_size));
@@ -138,10 +167,13 @@ __libc_setup_tls (size_t tcbsize, size_t tcbalign)
 
   /* Initialize the TLS block.  */
 # if TLS_TCB_AT_TP
-  static_dtv[2].pointer = tlsblock;
+  static_dtv[2].pointer = ((char *) tlsblock + tcb_offset
+			   - roundup (memsz, align));
+  static_map.l_tls_offset = roundup (memsz, align);
 # elif TLS_DTV_AT_TP
   tcb_offset = roundup (tcbsize, align);
   static_dtv[2].pointer = (char *) tlsblock + tcb_offset;
+  static_map.l_tls_offset = tcb_offset;
 # else
 #  error "Either TLS_TCB_AT_TP or TLS_DTV_AT_TP must be defined"
 # endif
@@ -152,8 +184,6 @@ __libc_setup_tls (size_t tcbsize, size_t tcbalign)
 
   /* Initialize the thread pointer.  */
 # if TLS_TCB_AT_TP
-  tcb_offset = roundup (memsz, tcbalign);
-
   INSTALL_DTV ((char *) tlsblock + tcb_offset, static_dtv);
 
   TLS_INIT_TP ((char *) tlsblock + tcb_offset, 0);
@@ -171,39 +201,38 @@ __libc_setup_tls (size_t tcbsize, size_t tcbalign)
   static_map.l_tls_blocksize = memsz;
   static_map.l_tls_initimage = initimage;
   static_map.l_tls_initimage_size = filesz;
-  static_map.l_tls_offset = tcb_offset;
   static_map.l_type = lt_executable;
   static_map.l_tls_modid = 1;
 
-  /* Create the slotinfo list.  */
-  static_slotinfo.si.len = (((char *) (&static_slotinfo + 1)
-			     - (char *) &static_slotinfo.si.slotinfo[0])
-			    / sizeof static_slotinfo.si.slotinfo[0]);
-  // static_slotinfo.si.next = NULL;	already zero
-
-  static_slotinfo.si.slotinfo[1].gen = 0;
+  init_slotinfo ();
+  // static_slotinfo.si.slotinfo[1].gen = 0; already zero
   static_slotinfo.si.slotinfo[1].map = &static_map;
 
-  /* The slotinfo list.  Will be extended by the code doing dynamic
-     linking.  */
-  GL(dl_tls_max_dtv_idx) = 1;
-  GL(dl_tls_dtv_slotinfo_list) = &static_slotinfo.si;
-
   memsz = roundup (memsz, align ?: 1);
+
 # if TLS_TCB_AT_TP
   memsz += tcbsize;
 # endif
 
-  /* That is the size of the TLS memory for this object.  The initialized
-     value of _dl_tls_static_size is provided by dl-open.c to request some
-     surplus that permits dynamic loading of modules with IE-model TLS.  */
-  GL(dl_tls_static_size) = roundup (memsz + GL(dl_tls_static_size),
-				    TLS_TCB_ALIGN);
-  GL(dl_tls_static_used) = memsz;
-  /* The alignment requirement for the static TLS block.  */
-  GL(dl_tls_static_align) = MAX (TLS_TCB_ALIGN, max_align);
-  /* Number of elements in the static TLS block.  */
-  GL(dl_tls_static_nelem) = GL(dl_tls_max_dtv_idx);
+  init_static_tls (memsz, MAX (TLS_TCB_ALIGN, max_align));
+}
+
+/* This is called only when the data structure setup was skipped at startup,
+   when there was no need for it then.  Now we have dynamically loaded
+   something needing TLS, or libpthread needs it.  */
+int
+internal_function
+_dl_tls_setup (void)
+{
+  init_slotinfo ();
+  init_static_tls (
+# if TLS_TCB_AT_TP
+		   TLS_TCB_SIZE,
+# else
+		   0,
+# endif
+		   TLS_TCB_ALIGN);
+  return 0;
 }
 
 
diff --git a/sysdeps/i386/i486/bits/string.h b/sysdeps/i386/i486/bits/string.h
index cc8594ff36..3a2ee2f6d7 100644
--- a/sysdeps/i386/i486/bits/string.h
+++ b/sysdeps/i386/i486/bits/string.h
@@ -636,9 +636,10 @@ __strcpy_g (char *__dest, __const char *__src)
      "leal	1(%1),%1\n\t"
      "testb	%b2,%b2\n\t"
      "jne	1b"
-     : "=&r" (__src), "=&r" (__tmp), "=&q" (__dummy)
+     : "=&r" (__src), "=&r" (__tmp), "=&q" (__dummy),
+       "=m" ( *(struct { char __x[0xfffffff]; } *)__dest)
      : "0" (__src), "1" (__tmp),
-       "m" ( *(struct { char __x[0xfffffff]; } *)__dest)
+       "m" ( *(struct { char __x[0xfffffff]; } *)__src)
      : "cc");
   return __dest;
 }