For variant II static TLS, properly align tls segments. Pre-calculate

the max required alignment for the static tls segments, and honor it
when carving the pieces for next module, from the static space.  Use
aligned allocator to get properly-aligned dynamic blocks.

Reported by:	dt71@gmx.com
Reviewed by:	kan
Sponsored by:	The FreeBSD Foundation
MFC after:	1 week

libexec/rtld-elf/rtld.c

@@ -231,6 +231,7 @@ char **main_argv;
 size_t tls_last_offset;		/* Static TLS offset of last module */
 size_t tls_last_size;		/* Static TLS size of last module */
 size_t tls_static_space;	/* Static TLS space allocated */
+size_t tls_static_max_align;
 int tls_dtv_generation = 1;	/* Used to detect when dtv size changes  */
 int tls_max_index = 1;		/* Largest module index allocated */
 
@@ -4281,19 +4282,22 @@ void *
 allocate_tls(Obj_Entry *objs, void *oldtls, size_t tcbsize, size_t tcbalign)
 {
     Obj_Entry *obj;
-    size_t size;
+    size_t size, ralign;
     char *tls;
     Elf_Addr *dtv, *olddtv;
     Elf_Addr segbase, oldsegbase, addr;
     int i;
 
-    size = round(tls_static_space, tcbalign);
+    ralign = tcbalign;
+    if (tls_static_max_align > ralign)
+	    ralign = tls_static_max_align;
+    size = round(tls_static_space, ralign) + round(tcbsize, ralign);
 
     assert(tcbsize >= 2*sizeof(Elf_Addr));
-    tls = xcalloc(1, size + tcbsize);
+    tls = malloc_aligned(size, ralign);
     dtv = xcalloc(tls_max_index + 2, sizeof(Elf_Addr));
 
-    segbase = (Elf_Addr)(tls + size);
+    segbase = (Elf_Addr)(tls + round(tls_static_space, ralign));
     ((Elf_Addr*)segbase)[0] = segbase;
     ((Elf_Addr*)segbase)[1] = (Elf_Addr) dtv;
 
@@ -4345,8 +4349,8 @@ allocate_tls(Obj_Entry *objs, void *oldtls, size_t tcbsize, size_t tcbalign)
 void
 free_tls(void *tls, size_t tcbsize, size_t tcbalign)
 {
-    size_t size;
     Elf_Addr* dtv;
+    size_t size, ralign;
     int dtvsize, i;
     Elf_Addr tlsstart, tlsend;
 
@@ -4354,19 +4358,22 @@ free_tls(void *tls, size_t tcbsize, size_t tcbalign)
      * Figure out the size of the initial TLS block so that we can
      * find stuff which ___tls_get_addr() allocated dynamically.
      */
-    size = round(tls_static_space, tcbalign);
+    ralign = tcbalign;
+    if (tls_static_max_align > ralign)
+	    ralign = tls_static_max_align;
+    size = round(tls_static_space, ralign);
 
     dtv = ((Elf_Addr**)tls)[1];
     dtvsize = dtv[1];
     tlsend = (Elf_Addr) tls;
     tlsstart = tlsend - size;
     for (i = 0; i < dtvsize; i++) {
-	if (dtv[i+2] && (dtv[i+2] < tlsstart || dtv[i+2] > tlsend)) {
+	if (dtv[i + 2] != 0 && (dtv[i + 2] < tlsstart || dtv[i + 2] > tlsend)) {
-	    free((void*) dtv[i+2]);
+		free_aligned(dtv[i + 2]);
 	}
     }
 
-    free((void*) tlsstart);
+    free_aligned(tlsstart);
     free((void*) dtv);
 }
 
@@ -4390,11 +4397,7 @@ allocate_module_tls(int index)
 	die();
     }
 
-    p = malloc(obj->tlssize);
+    p = malloc_aligned(obj->tlssize, obj->tlsalign);
-    if (p == NULL) {
-	_rtld_error("Cannot allocate TLS block for index %d", index);
-	die();
-    }
     memcpy(p, obj->tlsinit, obj->tlsinitsize);
     memset(p + obj->tlsinitsize, 0, obj->tlssize - obj->tlsinitsize);
 
@@ -4426,9 +4429,11 @@ allocate_tls_offset(Obj_Entry *obj)
      * leave a small amount of space spare to be used for dynamically
      * loading modules which use static TLS.
      */
-    if (tls_static_space) {
+    if (tls_static_space != 0) {
 	if (calculate_tls_end(off, obj->tlssize) > tls_static_space)
 	    return false;
+    } else if (obj->tlsalign > tls_static_max_align) {
+	    tls_static_max_align = obj->tlsalign;
     }
 
     tls_last_offset = obj->tlsoffset = off;